1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
|
$NetBSD: Debug.tips,v 1.2 1994/10/26 07:22:49 cgd Exp $
NOTE: this description applies to the hp300 system with the old BSD
virtual memory system. It has not been updated to reflect the new,
Mach-derived VM system, but should still be useful.
The new system has no fixed-address "u.", but has a fixed mapping
for the kernel stack at 0xfff00000.
--------------------------------------------------------------------------
Some quick notes on the HPBSD VM layout and kernel debugging.
Physical memory:
Physical memory always ends at the top of the 32 bit address space; i.e. the
last addressible byte is at 0xFFFFFFFF. Hence, the start of physical memory
varies depending on how much memory is installed. The kernel variable "lowram"
contains the starting locatation of memory as provided by the ROM.
The low 128k (I think) of the physical address space is occupied by the ROM.
This is accessible via /dev/mem *only* if the kernel is compiled with DEBUG.
[ Maybe it should always be accessible? ]
Virtual address spaces:
The hardware page size is 4096 bytes. The hardware uses a two-level lookup.
At the highest level is a one page segment table which maps a page table which
maps the address space. Each 4 byte segment table entry (described in
hp300/pte.h) contains the page number of a single page of 4 byte page table
entries. Each PTE maps a single page of address space. Hence, each STE maps
4Mb of address space and one page containing 1024 STEs is adequate to map the
entire 4Gb address space.
Both page and segment table entries look similar. Both have the page frame
in the upper part and control bits in the lower. This is the opposite of
the VAX. It is easy to convert the page frame number in an STE/PTE to a
physical address, simply mentally mask out the low 12 bits. For example
if a PTE contains 0xFF880019, the physical memory location mapped starts at
0xFF880000.
Kernel address space:
The kernel resides in its own virtual address space independent of all user
processes. When the processor is in supervisor mode (i.e. interrupt or
exception handling) it uses the kernel virtual mapping. The kernel segment
table is called Sysseg and is allocated statically in hp300/locore.s. The
kernel page table is called Systab is also allocated statically in
hp300/locore.s and consists of the usual assortment of SYSMAPs.
The size of Systab (Syssize) depends on the configured size of the various
maps but as currently configured is 9216 PTEs. Both segment and page tables
are initialized at bootup in hp300/locore.s. The segment table never changes
(except for bits maintained by the hardware). Portions of the page table
change as needed. The kernel is mapped into the address space starting at 0.
Theoretically, any address in the range 0 to Syssize * 4096 (0x2400000 as
currently configured) is valid. However, certain addresses are more common
in dumps than others. Those are (for the current configuration):
0 - 0x800000 kernel text and permanent data structures
0x917000 - 0x91a000 u-area; 1st page is user struct, last k-stack
0x1b1b000 - 0x2400000 user page tables, also kmem_alloc()ed data
User address space:
The user text and data are loaded starting at VA 0. The user's stack starts
at 0xFFF00000 and grows toward lower addresses. The pages above the user
stack are used by the kernel. From 0xFFF00000 to 0xFFF03000 is the u-area.
The 3 PTEs for this range map (read-only) the same memory as does 0x917000
to 0x91a000 in the kernel address space. This address range is never used
by the kernel, but exists for utilities that assume that the u-area sits
above the user stack. The pages from FFF03000 up are not used. They
exist so that the user stack is in the same location as in HPUX.
The user segment table is allocated along with the page tables from Usrptmap.
They are contiguous in kernel VA space with the page tables coming before
the segment table. Hence, a process has p_szpt+1 pages allocated starting
at kernel VA p_p0br.
The user segment table is typically very sparse since each entry maps 4Mb.
There are usually only two valid STEs, one at the start mapping the text/data
potion of the page table, and one at the end mapping the stack/u-area. For
example if the segment table was at 0xFFFFA000 there would be valid entries
at 0xFFFFA000 and 0xFFFFAFFC.
Random notes:
An important thing to note is that there are no hardware length registers
on the HP. This implies that we cannot "pack" data and stack PTEs into the
same page table page. Hence, every user page table has at least 2 pages
(3 if you count the segment table).
The HP maintains the p0br/p0lr and p1br/p1lr PCB fields the same as the
VAX even though they have no meaning to the hardware. This also keeps many
utilities happy.
There is no seperate interrupt stack (right now) on the HPs. Interrupt
processing is handled on the kernel stack of the "current" process.
Following is a list of things you might want to be able to do with a kernel
core dump. One thing you should always have is a ps listing from the core
file. Just do:
ps klaw vmunix.? vmcore.?
Exception related panics (i.e. those detected in hp300/trap.c) will dump
out various useful information before panicing. If available, you should
get this out of the /usr/adm/messages file. Finally, you should be in adb:
adb -k vmunix.? vmcore.?
Adb -k will allow you to examine the kernel address space more easily.
It automatically maps kernel VAs in the range 0 to 0x2400000 to physical
addresses. Since the kernel and user address spaces overlap (i.e. both
start at 0), adb can't let you examine the address space of the "current"
process as it does on the VAX.
--------
1. Find out what the current process was at the time of the crash:
If you have the dump info from /usr/adm/messages, it should contain the
PID of the active process. If you don't have this info you can just look
at location "Umap". This is the PTE for the first page of the u-area; i.e.
the user structure. Forget about the last 3 hex digits and compare the top
5 to the ADDR column in the ps listing.
2. Locating a process' user structure:
Get the ADDR field of the desired process from the ps listing. This is the
page frame number of the process' user structure. Tack 3 zeros on to the
end to get the physical address. Note that this doesn't give you the kernel
stack since it is in a different page than the user-structure and pages of
the u-area are not physically contiguous.
3. Locating a process' proc structure:
First find the process' user structure as described above. Find the u_procp
field at offset 0x200 from the beginning. This gives you the kernel VA of
the proc structure.
4. Locating a process' page table:
First find the process' user structure as described above. The first part
of the user structure is the PCB. The second longword (third field) of the
PCB is pcb_ustp, a pointer to the user segment table. This pointer is
actually the page frame number. Again adding 3 zeros yields the physical
address. You can now use the values in the segment table to locate the
page tables. For example, to locate the first page of the text/data part
of the page table, use the first STE (longword) in the segment table.
5. Locating a process' kernel stack:
First find the process' page table as described above. The kernel stack
is near the end of the user address space. So, locate the last entry in the
user segment table (base+0xFFC) and use that entry to find the last page of
the user page table. Look at the last 256 entries of this page
(pagebase+0xFE0) The first is the PTE for the user-structure. The second
was intended to be a read-only page to protect the user structure from the
kernel stack. Currently it is read/write and actually allocated. Hence
it can wind up being a second page for the kernel stack. The third is the
kernel stack. The last 253 should be zero. Hence, indirecing through the
third of these last 256 PTEs will give you the kernel stack page.
An alternate way to do this is to use the p_addr field of the proc structure
which is found as described above. The p_addr field is at offset 0x10 in the
proc structure and points to the first of the PTEs mentioned above (i.e. the
user structure PTE).
6. Interpreting the info in a "trap type N..." panic:
As mentioned, when the kernel crashes out of hp300/trap.c it will dump some
useful information. This dates back to the days when I was debugging the
exception handling code and had no kernel adb or even kernel crash dump code.
"trap type" (decimal) is as defined in hp300/trap.h, it doesn't really
correlate with anything useful. "code" (hex) is only useful for MMU
(trap type 8) errors. It is the concatination of the MMU status register
(see hp300/cpu.h) in the high 16 bits and the 68020 special status word
(see the 020 manual page 6-17) in the low 16. "v" (hex) is the virtual
address which caused the fault. "pid" (decimal) is the ID of the process
running at the time of the exception. Note that if we panic in an interrupt
routine, this process may not be related to the panic. "ps" (hex) is the
value of the 68020 status register (see page 1-4 of 020 manual) at the time
of the crash. If the 0x2000 bit is on, we were in supervisor (kernel) mode
at the time, otherwise we were in user mode. "pc" (hex) is the value of the
PC saved on the hardware exception frame. It may *not* be the PC of the
instruction causing the fault (see the 020 manual for details). The 0x2000
bit of "ps" dictates whether this is a kernel or user VA. "sfc" and "dfc"
are the 68020 source/destination function codes. They should always be one.
"p0" and "p1" are the VAX-like region registers. They are of the form:
<length> '@' <kernel VA>
where both are in hex. Following these values are a dump of the processor
registers (hex). Check the address registers for values close to "v", the
fault address. Most faults are causes by dereferences of bogus pointers.
Most such dereferences are the result of 020 instructions using the:
<address-register> '@' '(' offset ')'
addressing mode. This can help you track down the faulting instruction (since
the PC may not point to it). Note that the value of a7 (the stack pointer) is
ALWAYS the user SP. This is brain-dead I know. Finally, is a dump of the
stack (user/kernel) at the time of the offense. Before kernel crash dumps,
this was very useful.
7. Converting kernel virtual address to a physical address.
Adb -k already does this for you, but sometimes you want to know what the
resulting physical address is rather than what is there. Doing this is
simply a matter of indexing into the kernel page table. In theory we would
first have to do a lookup in the kernel segment table, but we know that the
kernel page table is physically contiguous so this isn't necessary. The
base of the system page table is "Sysmap", so to convert an address V just
divide the address by 4096 to get the page number, multiply that by 4 (the
size of a PTE in bytes) to get a byte offset, and add that to "Sysmap".
This gives you the address of the PTE mapping V. You can then get the
physical address by masking out the low 12 bits of the contents of that PTE.
To wit:
*(Sysmap+(VA%1000*4))&fffff000
where VA is the virtual address in question.
This technique should also work for user virtual addresses if you replace
"Sysmap" with the value of the appropriate processes' P0BR. This works
because a user's page table is *virtually* contiguous in the kernel
starting at P0BR, and adb will handle translating the kernel virtual addresses
for you.
|
