This is an old revision of the document!
==== Port wscale branch against current gPXE master ==== Recent TCP updates is a lot different from previous base of wscale branch. === Hit a GCC optimization trap === == For the following code == <code> #include <stdio.h> struct foo { char huge_gap[0x4005d0]; int v; }; void store_and_print(int *val, struct foo *f) { *val = f->v; if ( !f ) { printf("Value is wrong.\n"); } else { printf("Value is correct.\n"); } } int main() { int unidentified_value; printf("%p\n", main); store_and_print(&unidentified_value, NULL); return 0; } </code> The value ''0x4005d0'' is from ''printf("%p\n", main);'' in order to get an mapped memory address to prevent ''Segment fault'' error while running. == The output if compiled with ''-O2'' == <code> $ ./null_deref_optimized_away 0x4005d0 Value is correct. $ </code> And its assembly code: <code> void store_and_print(int *val, struct foo *f) { *val = f->v; 4005b0: 8b 86 d0 05 40 00 mov 0x4005d0(%rsi),%eax 4005b6: 89 07 mov %eax,(%rdi) if ( !f ) { printf("Value is wrong.\n"); } else { printf("Value is correct.\n"); 4005b8: bf ec 06 40 00 mov $0x4006ec,%edi 4005bd: e9 c6 fe ff ff jmpq 400488 <puts@plt> 4005c2: 66 66 66 66 66 2e 0f nopw %cs:0x0(%rax,%rax,1) 4005c9: 1f 84 00 00 00 00 00 00000000004005d0 <main>: } } </code> == The output if compiled with ''-O2 -fno-delete-null-pointer-checks'' == <code> $ ./null_deref_optimized_away 0x4005d0 Value is wrong. $ </code> And its assembly code: <code> 00000000004005b0 <store_and_print>: int v; }; void store_and_print(int *val, struct foo *f) { *val = f->v; 4005b0: 8b 86 d0 05 40 00 mov 0x4005d0(%rsi),%eax if ( !f ) { 4005b6: 48 85 f6 test %rsi,%rsi int v; }; void store_and_print(int *val, struct foo *f) { *val = f->v; 4005b9: 89 07 mov %eax,(%rdi) if ( !f ) { 4005bb: 74 13 je 4005d0 <store_and_print+0x20> printf("Value is wrong.\n"); } else { printf("Value is correct.\n"); 4005bd: bf 0c 07 40 00 mov $0x40070c,%edi 4005c2: e9 c1 fe ff ff jmpq 400488 <puts@plt> 4005c7: 66 0f 1f 84 00 00 00 nopw 0x0(%rax,%rax,1) 4005ce: 00 00 void store_and_print(int *val, struct foo *f) { *val = f->v; if ( !f ) { printf("Value is wrong.\n"); 4005d0: bf fc 06 40 00 mov $0x4006fc,%edi 4005d5: e9 ae fe ff ff jmpq 400488 <puts@plt> 4005da: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) 00000000004005e0 <main>: printf("Value is correct.\n"); } } </code> ==== Adding some facilities for testing convenience ==== * Add a command to modify TCP window size at runtime. * print "\b%c" instead of "." in monojob, keep the console clean. * Add a ''pause'' command. [[http://git.etherboot.org/?p=people/cooldavid/gpxe.git;a=shortlog;h=refs/heads/wscale2_test|wscale2 testing branch]] [[http://git.etherboot.org/?p=people/cooldavid/gpxe.git;a=shortlog;h=refs/heads/tcpfix2_test|tcpfix2 testing branch]] ==== Window size difference ==== * Test target image on BKO: http://130.239.17.17/bko/live/fedora/f13-alpha/i686/initrd0.img * Image size: 43,355,267 bytes * Linux wget speed: 1MByte/s (My Internet link speed is 10Mbit/s, it's about the full spped.) * Tested it with emulated rtl8139 of KVM, and Native on rtl8139 hardware. ^ Window Size ^ tcpfix2 branch ^^ wscale2 branch ^^ sackws2 branch ^^ ^ ^ KVM TAP/NAT ^ Native ^ KVM TAP/NAT ^ Native ^ KVM TAP/NAT ^ Native ^ | 8K| 17KB/s| 16KB/s| 16KB/s| 16KB/s| KB/s| KB/s| | 16K| 43KB/s| 43KB/s| 41KB/s| 41KB/s| KB/s| KB/s| | 32K| 94KB/s| 93KB/s| 90KB/s| 91KB/s| KB/s| KB/s| | 64K| 198KB/s| 192KB/s| 187KB/s| 187KB/s| KB/s| KB/s| | 128K| | | KB/s| 374KB/s| KB/s| KB/s| | 256K| | | KB/s| 742KB/s| KB/s| KB/s| | 512K| | | KB/s| [1]344KB/s| KB/s| KB/s| | 1024K| | | KB/s| [1]403KB/s| KB/s| KB/s| [1]: While TCP rx window is large enough to hit the bandwidth limit(1MB/s), packets start to drop. According to captured data, the retransmittion is very slow. The sending host wait a long time for transmit next missing packet required by client. This issue should be able to fixed with Selective ACK support.
