This reverts an earlier patch which disabled bidirectional mode, meaning that
a listening (passive) socket was not allowed to write to the other (active)
end of the connection.

This mode had been disabled when there were problems with CCID3, but it
imposes a constraint on socket programming and thus hinders deployment.

A change is included to ignore RX feedback received by the TX CCID3 module.

Many thanks to Andre Noll for pointing out this issue.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This mirrors a recent change in tcp_open_req_child, whereby the icsk_rto of the
newly created child socket was not set (but rather on the parent socket). Same
fix for DCCP.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This fixes a bug caused by a previous patch, which causes DCCP servers in
LISTEN state to not receive packets.

This patch changes the logic so that
 * servers in either LISTEN or OPEN state get the RX half connection packets
 * clients in OPEN state get the TX half connection packets
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NPatrick McHardy <kaber@trash.net>
Acked-by: NPaul Moore <paul.moore@hp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The semantic effect of insert_at_head is that it would allow new registered
sysctl entries to override existing sysctl entries of the same name.  Which is
pain for caching and the proc interface never implemented.

I have done an audit and discovered that none of the current users of
register_sysctl care as (excpet for directories) they do not register
duplicate sysctl entries.

So this patch simply removes the support for overriding existing entries in
the sys_sysctl interface since no one uses it or cares and it makes future
enhancments harder.
Signed-off-by: NEric W. Biederman <ebiederm@xmission.com>
Acked-by: NRalf Baechle <ralf@linux-mips.org>
Acked-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: David Howells <dhowells@redhat.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Andi Kleen <ak@muc.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Corey Minyard <minyard@acm.org>
Cc: Neil Brown <neilb@suse.de>
Cc: "John W. Linville" <linville@tuxdriver.com>
Cc: James Bottomley <James.Bottomley@steeleye.com>
Cc: Jan Kara <jack@ucw.cz>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Cc: David Chinner <dgc@sgi.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Patrick McHardy <kaber@trash.net>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: NEric W. Biederman <ebiederm@xmission.com>
Cc: Arnaldo Carvalho de Melo <acme@conectiva.com.br>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Many struct file_operations in the kernel can be "const".  Marking them const
moves these to the .rodata section, which avoids false sharing with potential
dirty data.  In addition it'll catch accidental writes at compile time to
these shared resources.
Signed-off-by: NArjan van de Ven <arjan@linux.intel.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: NYOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ehash table layout is currently this one :

First half of this table is used by sockets not in TIME_WAIT state
Second half of it is used by sockets in TIME_WAIT state.

This is non optimal because of for a given hash or socket, the two chain heads 
are located in separate cache lines.
Moreover the locks of the second half are never used.

If instead of this halving, we use two list heads in inet_ehash_bucket instead 
of only one, we probably can avoid one cache miss, and reduce ram usage, 
particularly if sizeof(rwlock_t) is big (various CONFIG_DEBUG_SPINLOCK, 
CONFIG_DEBUG_LOCK_ALLOC settings). So we still halves the table but we keep 
together related chains to speedup lookups and socket state change.

In this patch I did not try to align struct inet_ehash_bucket, but a future 
patch could try to make this structure have a convenient size (a power of two 
or a multiple of L1_CACHE_SIZE).
I guess rwlock will just vanish as soon as RCU is plugged into ehash :) , so 
maybe we dont need to scratch our heads to align the bucket...

Note : In case struct inet_ehash_bucket is not a power of two, we could 
probably change alloc_large_system_hash() (in case it use __get_free_pages()) 
to free the unused space. It currently allocates a big zone, but the last 
quarter of it could be freed. Again, this should be a temporary 'problem'.

Patch tested on ipv4 tcp only, but should be OK for IPV6 and DCCP.
Signed-off-by: NEric Dumazet <dada1@cosmosbay.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

net/dccp/ccids/ccid3.c: In function `ccid3_hc_rx_packet_recv':
net/dccp/ccids/ccid3.c:1007: warning: long int format, different type arg (arg 3)
net/dccp/ccids/ccid3.c:1007: warning: long int format, different type arg (arg 4)

opaque types must be suitably cast for printing.
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Do this even for non-blocking sockets.  This avoids the silly -EAGAIN
that applications can see now, even for non-blocking sockets in some
cases (f.e. connect()).

With help from Venkat Tekkirala.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Revert 93173112

We can't elide the skb_set_owner_w() here because things like certain
netfilter targets (such as owner MATCH) need a socket to be set on the
SKB for correct operation.

Thanks to Jan Engelhardt and other netfilter list members for
pointing this out.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In a recent patch we introduced invalid return codes which will result in the
opposite of what is intended (i.e. send more packets in face of peculiar
network conditions).

This fixes it by returning ~0 which means not calculated as per
dccp_li_hist_calc_i_mean.
Signed-off-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

That accumulated over the last months hackaton, shame on me for not
using git-apply whitespace helping hand, will do that from now on.
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

Spotted by David Miller when compiling on sparc64, I reproduced it here on
parisc64, that are the only platforms to define __kernel_suseconds_t as an
'int', all the others, x86_64 and x86 included typedef it as a 'long', but from
the definition of suseconds_t it should just be an 'int' on platforms where it
is >= 32bits, it would not require all the castings from suseconds_t to (int)
when printking variables of this type, that are not needed on parisc64 and
sparc64.
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This fixes conversion errors which arose by not properly type-casting
from u32 to __u64. Fixed by explicitly casting each type which is not
__u64, or by performing operation after assignment.

The patch further adds missing debug information to track the current
value of X_recv.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

No code change at all.

This reorders the source file to follow the same order as the corresponding
header file.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

No code change at all.

To make the header file easier to read, the following ordering is established
among the declarations:
	* hist_new
	* hist_delete
	* hist_entry_new
	* hist_head
	* hist_find_entry
	* hist_add_entry
	* hist_entry_delete
	* hist_purge
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This patch does not alter any algorithm, just the debug message format:

 * s#%s, sk=%p#%s(%p)#g

 * when a statename is present, it now uses %s(%p, state=%s)

 * when only function entry is debugged, it adds an `- entry'
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

 This migrates all packet history operations into the routine
 ccid3_hc_tx_packet_sent, thereby removing synchronization problems
 that occur when, as before, the operations are spread over multiple
 routines.
 The following minor simplifications are also applied:
  * several simplifications now follow from this change - several tests
    are now no longer required
  * removal of one unnecessary variable (dp)

Justification:

 Currently packet history operations span two different routines,
 one of which is likely to pass through several iterations of sleeping
 and awakening.
 The first routine, ccid3_hc_tx_send_packet, allocates an entry and
 sets a few fields. The remaining fields are filled in when the second
 routine (which is not within a sleeping context), ccid3_hc_tx_packet_sent,
 is called. This has several strong drawbacks:
  * it is not necessary to split history operations - all fields can be
    filled in by the second routine
  * the first routine is called multiple times, until a packet can be sent,
    and sleeps meanwhile - this causes a lot of difficulties with regard to
    keeping the list consistent
  * since both routines do not have a producer-consumer like synchronization,
    it is very difficult to maintain data across calls to these routines
  * the fact that the routines are called in different contexts (sleeping, not
    sleeping) adds further problems
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This removes the `dccphtx_ccval' field since it is nowhere used in the code and
in fact not necessary for the accounting.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

 This puts the window counter computation [RFC 4342, 8.1] into a separate
 function which is called whenever a new packet is ready for immediate
 transmission in ccid3_hc_tx_send_packet.

Justification:

 The window counter update was previously computed after the packet was sent. This has
 two drawbacks, both fixed by this patch:
   1) re-compute another timestamp almost directly after the packet was sent (expensive),
   2) the CCVal for the window counter is needed at the instant the packet is sent.

Further details:

 The initialisation of the window counter is left in the state NO_SENT, as before.
 The algorithm will do nothing if either RTT is initialised to 0 (which is ok) or if
 the RTT value remains below 4 microseconds (which is almost pathological).
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

CCID3 performance depends much on the accuracy of RTT samples.  If RTT
samples grow too large, performance can be catastrophically poor.

To limit the amount of possible damage in such cases, the patch
 * introduces an upper limit which identifies a maximum `sane' RTT value;
 * uses a macro to enforce this upper limit.

Using a macro was given preference, since it is necessary to identify the
calling function in the warning message. Since exceeding this threshold
identifies a critical condition, DCCP_CRIT is used and not DCCP_WARN.

Many thanks to Ian McDonald for collaboration on this issue.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

In both the sender and the receiver it is possible that the stored
RTT value is accessed before an actual RTT estimate has been computed.

This patch
 * initialises the sender RTT to 0
     - the sender always accesses the RTT in ccid3_hc_tx_packet_sent
     - the RTT is further needed for the window counter algorithm

 * replaces the receiver initialisation of 5msec with 0
     - which has the same effect and removes an `XXX'
     - the RTT value is needed in ccid3_hc_rx_packet_recv as rtt_prev
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

 The function ccid3_hc_tx_insert_options only does a redundant no-op,
 as the operation

  DCCP_SKB_CB(skb)->dccpd_ccval = hctx->ccid3hctx_last_win_count;

 is already performed _unconditionally_ in ccid3_hc_tx_send_packet.

 Since there is further no current need for this function, it is removed
 entirely. Since furthermore, there is actually no present need for the
 entire interface function ccid_hc_tx_insert_options, it was decided to
 remove it also, to clean up the interface.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This adds a (debug) warning message which is triggered whenever a packet is
discarded due to send failure.

It also adds a conditional, so that an interruption during dccp_wait_for_ccid
is not treated as a `BUG': the rationale is that interruptions are external,
whereas bug warnings are concerned with the internals.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This is an optimisation to reduce CPU load. The received feedback is now
only directed to the active CCID component, without requiring processing
also by the inactive one.

As a consequence, a similar test in ccid3.c is now redundant and is
also removed.

Justification:

 Currently DCCP works as a unidirectional service, i.e. a listening server
 is not at the same time a connecting client.
 As far as I can see, several modifications are necessary until that
 becomes possible.
 At the present time, received feedback is both fed to the rx/tx CCID
 modules. In unidirectional service, only one of these is active at any
 one time.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

In migrating towards using the newer functions scaled_div/scaled_div32
for TFRC computations mapped from floating-point onto integer arithmetic,
this completes the last stage of modifications.

In particular, the overflow case for computing X_calc is circumvented by
 * breaking the computation into two stages
 * the first stage, res = (s*1E6)/R, cannot overflow due to use of u64
 * in the second stage, res = (res*1E6)/f, overflow on u32 is avoided due
   to (i) returning UINT_MAX in this case (which is logically appropriate)
   and (ii) issuing a warning message into the system log (since very likely
   there is a problem somewhere else with the parameters)

Lastly, all such scaling operations are now exported into tfrc.h, since
actually this form of scaled computation is specific to TFRC and not to CCID3.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

Problem:

 Most target types in the CCID3 code are u32, so subtle conversion errors
 can occur if signed time calculations yield negative results: the original
 values are lost in the conversion to unsigned, calculation errors go undetected.

 This patch therefore
   * sets all critical time types from unsigned to suseconds_t
   * avoids comparison between signed/unsigned via type-casting
   * provides ample warning messages in case time calculations are negative

 These warning messages can be removed at a later stage when the code
 has undergone more testing.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

 This simplifies the calculation of a value p for a given fval when the
 first loss interval is computed (RFC 3448, 6.3.1). It makes use of the
 two new functions scaled_div/scaled_div32 to provide overflow protection.

 Additionally, protection against divide-by-zero is extended - in this
 case the function will return the maximally possible value of p=100%.

Background:

 The maximum fval, f(100%), is approximately 244, i.e. the scaled value of fval
 should never exceed 244E6, which fits easily into u32. The problem is the scaling
 by 10^6, since additionally R(TT) is in microseconds.
 This is resolved by breaking the division into two stages: the first stage
 computes fval=(s*10^6)/R, stores that into u64; the second stage computes
 fval = (fval*10^6)/X_recv and complains if overflow is reached for u32.
 This case is safe since the TFRC reverse-lookup routine then returns p=100%.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This replaces the remaining uses of usecs_div with scaled_div32, which
internally uses 64bit division and produces a warning on overflow.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This patch
 * resolves a bug where packets smaller than 32/64 bytes resulted in sending rates of 0
 * supports all sending rates from 1/64 bytes/second up to 4Gbyte/second
 * simplifies the present overflow problems in calculations

Current sending rate X and the cached value X_recv of the receiver-estimated
sending rate are both scaled by 64 (2^6) in order to
 * cope with low sending rates (minimally 1 byte/second)
 * allow upgrading to use a packets-per-second implementation of CCID 3
 * avoid calculation errors due to integer arithmetic cut-off

The patch implements a revised strategy from
http://www.mail-archive.com/dccp@vger.kernel.org/msg01040.html

The only difference with regard to that strategy is that t_ipi is already
used in the calculation of the nofeedback timeout, which saves one division.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This fixes
 1) a bug in the recomputation of the sending rate by the nofeedback
    timer when no feedback at all has so far been sent by the receiver:
    min_t was used instead of max_t, which is wrong (cf. RFC 3448, p. 10);

 2) an error in the computation of larger initial windows: instead of
    min(... max()) (cf. RFC 4342, 5.), the code had used max(... max()).
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This performs two optimisations for the recomputation of the sending rate.

1) Currently the target sending rate X_calc is recalculated whenever
	a) the nofeedback timer expires, or
	b) a feedback packet is received.
   In the (a) case, recomputing X_calc is redundant, since

    * the parameters p and RTT do not change in between the
      reception of feedback packets;

    * the parameter X_recv is either modified from received
      feedback or via the nofeedback timer;

    * a test (`p == 0') in the nofeedback timer avoids using
      a stale/undefined value of X_calc if p was previously 0.

2) The nofeedback timer now only recomputes a timestamp when p == 0.
   This is according to step (4) of [RFC 3448, 4.3] and avoids
   unnecessarily determining a timestamp.

A debug statement about not updating X is also removed - it helps very
little in debugging and just clutters the logs.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

This patch follows a suggestion by Ian McDonald and ensures that in
the current code the value of p can not exceed 100%.  Such a value is
illegal and would consequently cause a bug condition in tfrc_calc_x().

The receiver case is also tested, and a warning message is added.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

It simplifies waiting for the CCID module to signal that a packet
is ready to be sent.  Other simplifications flow on from this such as
removing constants.

As a result of this EAGAIN is not returned any more by dccp_wait_for_ccid
(which would otherwise lead to unnecessarily discarding the packet in
dccp_write_xmit).
Signed-off-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

Replace all uses of kmem_cache_t with struct kmem_cache.

The patch was generated using the following script:

	#!/bin/sh
	#
	# Replace one string by another in all the kernel sources.
	#

	set -e

	for file in `find * -name "*.c" -o -name "*.h"|xargs grep -l $1`; do
		quilt add $file
		sed -e "1,\$s/$1/$2/g" $file >/tmp/$$
		mv /tmp/$$ $file
		quilt refresh
	done

The script was run like this

	sh replace kmem_cache_t "struct kmem_cache"
Signed-off-by: NChristoph Lameter <clameter@sgi.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

SLAB_ATOMIC is an alias of GFP_ATOMIC
Signed-off-by: NChristoph Lameter <clameter@sgi.com>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This replaces the linear search algorithm for reverse lookup with
binary search.

It has the advantage of better scalability: O(log2(N)) instead of O(N).
This means that the average number of iterations is reduced from 250
(linear search if each value appears equally likely) down to at most 9.
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>

 This patch deprecates the existing use of an arbitrary value TFRC_SMALLEST_P
 for low-threshold values of p. This avoids masking low-resolution errors.
 Instead, the code now checks against real boundaries (implemented by preceding
 patch) and provides warnings whenever a real value falls below the threshold.

 If such messages are observed, it is a better solution to take this as an
 indication that the lookup table needs to be re-engineered.

Changelog:
----------
 This patch
   * makes handling all TFRC resolution errors local to the TFRC library

   * removes unnecessary test whether X_calc is 'infinity' due to p==0 -- this
     condition is already caught by tfrc_calc_x()

   * removes setting ccid3hctx_p = TFRC_SMALLEST_P in ccid3_hc_tx_packet_recv
     since this is now done by the TFRC library

   * updates BUG_ON test in ccid3_hc_tx_no_feedback_timer to take into account
     that p now is either 0 (and then X_calc is irrelevant), or it is > 0; since
     the handling of TFRC_SMALLEST_P is now taken care of in the tfrc library

Justification:
--------------
 The TFRC code uses a lookup table which has a bounded resolution.
 The lowest possible value of the loss event rate `p' which can be
 resolved is currently 0.0001.  Substituting this lower threshold for
 p when p is less than 0.0001 results in a huge, exponentially-growing
 error.  The error can be computed by the following formula:

    (f(0.0001) - f(p))/f(p) * 100      for p < 0.0001

 Currently the solution is to use an (arbitrary) value
     TFRC_SMALLEST_P  =   40 * 1E-6   =   0.00004
 and to consider all values below this value as `virtually zero'.  Due to
 the exponentially growing resolution error, this is not a good idea, since
 it hides the fact that the table can not resolve practically occurring cases.
 Already at p == TFRC_SMALLEST_P, the error is as high as 58.19%!
Signed-off-by: NGerrit Renker <gerrit@erg.abdn.ac.uk>
Acked-by: NIan McDonald <ian.mcdonald@jandi.co.nz>
Signed-off-by: NArnaldo Carvalho de Melo <acme@mandriva.com>