[realtek] rtl8139.c and 8139too.c old _BUG_

[Roman]

Hi ALL!!
I have (many) Acorp NICs with RTL8139B chip on many computers
(k6-2 266 on ali motherboard, celeron 366 on via motherboard ).
I've tested 8139too.c (v0.9.15, v0.9.17, v0.9.18, v0.9.22a) and
rtl8139.c (v1.12a, v1.13, v1.14, v1.15, v1.16a) and _ALL_ of them have
same bug (rtl8139.c v1.07 is good; I didn't get v.1.08 - v.1.11 because
I don't know where) with kernels 2.2.16, 2.2.19, 2.4.5, 2.4.17 as
modules and compiled into kernel, in 10T and 100TX modes.

THE BUG IS:
Over 20% of packets are lost and speed in 100Mbit/s mode was just
~ 900 Kbytes/s.
how I can see it (in 10T mode):
% ping 192.168.11.1 -s 1492
PING 192.168.11.7 (192.168.11.7): 1492 octets data
1500 octets from 192.168.11.7: icmp_seq=0 ttl=128 time=4.1 ms
1500 octets from 192.168.11.7: icmp_seq=1 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=2 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=3 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=4 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=5 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=6 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=9 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=10 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=13 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=14 ttl=128 time=3.9 ms
1500 octets from 192.168.11.7: icmp_seq=17 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=19 ttl=128 time=3.9 ms
1500 octets from 192.168.11.7: icmp_seq=20 ttl=128 time=3.8 ms
1500 octets from 192.168.11.7: icmp_seq=23 ttl=128 time=3.5 ms

--- 192.168.11.7 ping statistics ---
24 packets transmitted, 15 packets received, 37% packet loss

Please tell me where is problem (RX,TX rings change was tried).
May be problem in MMIO?
Where I can get versions 1.08-1.11 of rtl8139.c? (then I can locate
bug).

LOT OF THANKS FOR ANSWER.
-- 
Best regards,
 Roman                          mailto:roman@complex.lviv.ua