Timo Kokkonen [Tue, 29 Nov 2022 13:38:12 +0000 (15:38 +0200)]
Show full process argument list instead only executable name
Some systems have same executable running multiple times, executed
with different command line arguments. These may be difficult to
distinguish from each other if we don't see also the arguments.
Show the whole argument list for each processes as this usually what
we want to see.
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Tue, 29 Nov 2022 13:28:08 +0000 (15:28 +0200)]
parse: Fix out of bound access
Make sure the copy length is at least one shorter than the buffer
size. This isn't strictly required as we would be forcefully NULL
terminating the string aynway, but this silences the compiler warning
we are getting with modern versions of gcc.
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Sat, 29 Dec 2018 15:06:11 +0000 (17:06 +0200)]
Reverse list printing order
The process and mapping lists are often very long and rarely fit
entirely on the screen. Usually we are most interested in the largest
entries, the ones that now get printed in the beginning of the
list. This is cumbersome as it therefore requires quite a bit of
scrolling to see what you really want.
Reverse the list printing order so that the tool becomes more friendly
to the end user.
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Thu, 12 Jul 2012 14:41:49 +0000 (17:41 +0300)]
parse_maps: Fix issues with white space in file names
If there is any white space character in the file name field, sscanf
would only copy the file name part until the first white space and
leave out everything else.
The fix is trivial; we will use the %n with sscanf to find out all the
chars before the file name part and then skip that part when using
strncpy to copy the file name.
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Mon, 11 Jun 2012 19:19:55 +0000 (22:19 +0300)]
print_page_stats: Reserve 7 digits for page count alignment
If there are mappings that are larger than 4 gigabytes, 6 digits is
not enough for the output. 7 digits is enough until 40 gigabytes,
which should probably be enough in most of users. If someone has got
more than 40 gigabytes of ram and wants to have nicely aligned page
usage prints, I'm happy to implement some sort of dynamic
alignmenting for it..
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Mon, 11 Jun 2012 15:32:10 +0000 (18:32 +0300)]
pf_insert: Return matching node on insert
When a matching entry is found from the rb_tree, the insert function
should return the found entry from the tree to the caller. This is
also explicitly documented in the code. However, the code incorrectly
returned NULL instead.
Fix the code to return the match as it should do. Without this fix
identical pageframe entries are never located.
Signed-off-by: Timo Kokkonen <timo.t.kokkonen@iki.fi>
Timo Kokkonen [Mon, 7 Nov 2011 16:24:50 +0000 (18:24 +0200)]
update_kpageflags: Ensure file descriptors are closed correctly on error
If only opening only other of the files fails, the successfully opened
file was not closed. This scenario is extremely unlikely, but should
be fixed never the less.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 3 Nov 2011 12:54:41 +0000 (14:54 +0200)]
Fix bug when showing process shared mappings
The way page table tree is walked through nowadays requires that the
process map data is incremented every time the page table tree is
processed. Otherwise the last time the tree is processed, the counters
are overwritten with the last values, which is most likely zeros.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 16 Oct 2011 18:20:20 +0000 (21:20 +0300)]
print_pid_stats: Revert order of pageframe tree and process list are traversed
The pageframe tree is usually quite big, in order of dozens of
megabytes. The process list on the other hand is usually quite small,
usually in the order of tens of kilobytes. The process list thus can
fit entirely in L2 (or even in L1) cache whereas the pageframe tree
most likely does not fit in any cache at all.
Therefore, when calculating the pid statistics, walking through the
page frame tree as many times as there are processes in the process
list will likely cause very big traffic to the memory bus. Changing
the ordering of the loops will significantly increase cache locality
and reduce the memory bus traffic, decreasing the execution times a
lot.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sat, 15 Oct 2011 18:02:26 +0000 (21:02 +0300)]
Replace bintree with rbtree implementation
The biggest flaw with the bintree is that they can easily become
unbalanced. For example, if X server maps the video card memory into
its address space as 1:1 mapping, running scan-pagemap can be really
slow. This is because reading the 1:1 mapping will generate horribly
skewed binary tree that resembles more closely a single direction
linked list. Replacing the simple binary tree implementation with the
red black tree can make full scan over five times faster.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 2 Oct 2011 09:10:38 +0000 (12:10 +0300)]
Print verbose output during operation
Sometimes (especially with slower computers with a lot of memory, or
just computers with *a lot* of memory) it can take quite a while until
all the processing is done. Print brief status notes to the user
stating which pid's data we are now processing. This allows the user
to know how far we are going.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 29 Sep 2011 17:58:09 +0000 (20:58 +0300)]
main: Set a reasonable default function if no arguments are given
A very useful way to use scan-pagemap is to simply print out memory
usage of all processes on the system. Let this be the default behavior
of the program if no other options are given.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Mon, 6 Dec 2010 09:31:54 +0000 (11:31 +0200)]
Makefile: Reduce wildcard usage
Having wildcards in the Makefile is error prone in situations where
there are unrelated files in the source tree that are not related to
the actual project. For example, if some source files have been
renamed and there are leftover .*.o.d files that still contain the old
source files. Then make will incorrectly determine that some old files
would still be needed even though they are not part of the project any
more.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 2 Sep 2010 15:35:55 +0000 (18:35 +0300)]
main: Assume non-option arguments to be pids or process names
If arguments are given without any option arguments, it is tested if
it is a valid pid number, and add it to pidlist. If it is not a pid
number, then it is threated as a process name the system is scanned
for processes with a matching name.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 29 Aug 2010 10:30:33 +0000 (13:30 +0300)]
Analyzer: Print detailed kpageflag statistics
The print_page_statst() function is modified so that it will count how
many pages are found with each of the kpageflag bit set. The collected
stats are then printed so that count for every non-zero kpageflag is
shown. Since there are so many kpageflag flags (22 at the moment),
zero number counts are alwaysy omitted.
Since root priviledges are required to read the /proc/kpageflag
interface, all of the kpageflag counts are zero if user does not have
root priviledges.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 26 Aug 2010 18:43:32 +0000 (21:43 +0300)]
parser: Remove unaccessible pids from the pidlist
If scan-pagemap is run without root priviledges and --shared-mappings
option is used, pids that are inaccessible to the user need to be
removed from the list of interesting pids. Otherwise
dunp_process_maps() will not print any pages for the given processes,
as it will try to count pages that are shared with all interesting
pids. If any pids from the list are inaccessible to the user, then
none of the pages are shared with these inaccessible pids. Therefore,
the inaccessible pids must be removed from the list altogether.
Since list entries are being removed from the list, all
list_for_each_entry will be need to be replaced with the _safe
variant.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sat, 21 Aug 2010 07:24:07 +0000 (10:24 +0300)]
main.c: Allow scanning multiple processes by name
Now that we have code in place for scanning multiple pid numbers, we
can also convert all process names to pids. That way, it becomes also
possible to add all process names user has given as an argument.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sat, 21 Aug 2010 07:18:24 +0000 (10:18 +0300)]
Analyzer: Drop PRETTY_THRESH value to 98
We have reserved five digits for the byte sizes in the print
output. However, if the size of a process happens to be for example
101272kB, the size is still below the 100 * 1024 = 102400 kB
threshold, which has in fact six digits.
Lowering the threshold to 98 will ensure the pretty output is never
more than five digits.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Wed, 18 Aug 2010 16:34:19 +0000 (19:34 +0300)]
Mark initially scanned pids in analyzer printouts
Since the initial page mappings are are collected from one or more
pids, they deserve to be marked on the analyzer output. That way the
user will know which are the processes that have all of their pages
counted and which have only those pages counted that are shared with
any of the interesting processes.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Wed, 18 Aug 2010 16:31:13 +0000 (19:31 +0300)]
Add macro for testing parse option flags
The parse option structure parse_mask variable is tested for option
bits in quite many places. Add a macro that makes testing easier and
moves the duplicate code in one macro.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Wed, 18 Aug 2010 16:12:24 +0000 (19:12 +0300)]
Analyzer: Show prints in kilobyte resolution
Since all pages are 4kB in size, all mapping and process sizes are
also always divizable by 4kB. Therefore it makes no sense to show
sizes less than 100kB in bytes. Using kilobyte as a smallest possible
size improves readability.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Mon, 16 Aug 2010 18:23:30 +0000 (21:23 +0300)]
Compact pageframe structure
Instead of decoding each entry to individual variables in the
pageframe structure, maintain them as a single raw 64 bit
value. Various helper functions are introduced for decoding the
information on the fly. This reduces the run time memory consumption
without really slowing down the execution.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 15 Aug 2010 19:44:14 +0000 (22:44 +0300)]
print_page_stats: Print unique pages
Now that we don't scan interesting processes multiple times, we can
count the pages which have refcount == 1 as unique pages and show that
number to the user.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 15 Aug 2010 19:37:16 +0000 (22:37 +0300)]
parser: When scannin by name, scan all pids twice
When scanning by a mapping name, we must first scan through all
processes to find out which processes are using the mappings. Then we
need to scan all pids again to find out if there are somewhere
physical pages that are being shared by the given mapping with a
different name.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 15 Aug 2010 17:45:57 +0000 (20:45 +0300)]
Correct the usage of linked lists
There is really no documentation about how the linked lists in the
lists.h should be used. Specifically, it does not mention the fact
that the head entry of the list should never be used at all. Instead,
the head entry, when being empty, just points to itself.
Now that this fact is clear, a lot of difficult assumptions can be
left out and some overly complicated code can be cleared with the list
handling. This is because the head entry does not have to contain a
valid entry at all.
This also fixes a bug with was caused by the incorrect list handling:
printing the process list always had the head entry missing, since the
list_for_all() functions left the head entry unhandled for purpose.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 15 Aug 2010 16:43:08 +0000 (19:43 +0300)]
parser: Do not scan processes twice
Before we would scan some processes twice: First when we find out the
pages of the interesting processes, then again when we scan the rest
of the processes. This patch changes the handling so that the second
time we ignore all interesting pids that we already scanned.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sun, 15 Aug 2010 16:39:10 +0000 (19:39 +0300)]
read_pageframe: Do not add empty entries in process list
If reading of some of the process details fails, the process is most
likely died while we were processing its details. Thus, there is no
good to add the entry to the process list either. The entry is now
freed instead of adding the half-read entry to the list.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Sat, 14 Aug 2010 16:56:09 +0000 (19:56 +0300)]
parser: Fix the order of argumets for list_add()
The parameters were in wrong order, thus causing the new pageframe
referencing node to be lost for most of the processes. The results
were thus showing only references for only the first pid that was
found, ignoring everything else. This fix will thus improve the
results dramatically.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Fri, 23 Jul 2010 12:38:59 +0000 (15:38 +0300)]
Break parsing if no processes are found
Scanning through all pids in the system can be a time consuming task,
especially with slower CPUs. Therefore, if the user gives a
nonexistent process argument it can tage significant amount of time
until the user gets the response that there will be no useful results
available.
This patch counts the number of processes that have been encoutered
during the prescan. If the count is zero, then it is no good to
proceed scanning all of the rest processes. As an extra benefit, the
number of prescanned processes can be used for something useful..
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 22 Jul 2010 11:19:45 +0000 (14:19 +0300)]
parser: get_next_pid: Do not store state internally
Storing the directory state internally may conflict in case there are
multiple users to the get_next_pid function. The directory state is
now taken as an argument and caller will store it.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
Timo Kokkonen [Thu, 22 Jul 2010 09:12:07 +0000 (12:12 +0300)]
Scan threads only when requested
Usually there is no need to scan every threads for every processes, as
they are very likely to share their address space between each
others. Thus, scanning only one of them is mostly enough.
Signed-off-by: Timo Kokkonen <kaapeli@itanic.dy.fi>
projects / scan-pagemap / log