In the link referenced below there's an explanation of
the linux autogroup feature which has been around ever
since linux-2.6.38. With that explanation there's also
surprising (maybe shocking) revelations about the nice
and renice commands if CONFIG_SCHED_AUTOGROUP was set.
When autogroups are active, such programs are rendered
mostly useless because the nice value will only affect
scheduling priority relative to other processes in the
same autogroup. In order to accomplish what we thought
of as renice, that nice value in /proc/<pid>/autogroup
must be changed. Altering any single member of a group
will also affect every other member of that autogroup.
So, this commit will set the stage for users of newlib
to display autogroup identifiers plus their associated
nice values (now that their importance is understood).
Reference(s):
https://github.com/nlburgin/reallynice
Signed-off-by: Jim Warner <james.warner@comcast.net>
That recent addition of USS to the ps program prompted
this change. Rather than have it (and soon top) add 2
separate items to yield the desired value, we will let
our new library perform the arithmetic when necessary.
Outside of a little extra storage, there is no runtime
costs for such an extension. There is, however, a real
benefit to having such code in the library. Now should
callers choose to sort on this new field, results will
be guaranteed to be what was expected (i.e. accurate).
Signed-off-by: Jim Warner <james.warner@comcast.net>
The two special hugetlbfs items were misnamed. The TBL
reference (table) should be TLB (transaction lookaside
buffer). Besides, I never liked their position anyway!
[ and one macro argument tweak is being snuck in too ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
A couple of people have suggested that smaps_rollup be
added to the ps program and/or top program. This patch
is intended to set the stage for just such extensions.
There are currently 20 displayable items in the rollup
file. And newlib sometimes uses sscanf when populating
the target, sometimes hsearch and one customized gperf
approach. None of these fit well with the smaps items.
Thus, an approach using a simple table lookup was used
and, by disabling 1 code line, it could be made immune
from changes to the items order (unlike a sscanf call)
and doesn't carry the greater cost of a hsearch/gperf.
Note: The next patch will allow top to display some of
these new fields. Then, it'll be possible to determine
the colossal costs of accessing the smaps_rollup file.
Here is a small preview of just what you will discover
when using the command 'time top/top -d0 -n1000' while
configured with just two fields: PID + 1 memory field.
------------------------------------ as a regular user
with only PID + RES (statm)
real 0m2.605s
user 0m1.060s
sys 0m1.377s
with only PID + RSS (smaps)
real 0m26.397s 10x more costly
user 0m1.253s
sys 0m24.915s
----------------- as a root (thus smaps for all tasks)
with only PID + RES (statm)
real 0m2.651s
user 0m1.177s
sys 0m1.286s
with only PID + RSS (smaps)
real 0m33.040s 12x more costly
user 0m1.256s
sys 0m31.533s
Reference(s):
. ps: expose shared/private memory separately
https://gitlab.com/procps-ng/procps/-/issues/201
. top/ps: add support for PSS reporting
https://gitlab.com/procps-ng/procps/-/issues/112
Signed-off-by: Jim Warner <james.warner@comcast.net>
This commit is strictly cosmetic. It was an attempt to
normalize/standardize/alphabetize those #define/#undef
statements. Some missing #undef's were added plus some
comments regarding sources corrected/standardized too.
Signed-off-by: Jim Warner <james.warner@comcast.net>
This is a modification of MR !122 by @renit1609 to fit the new
library.
Problem statement:
The procps library has no PROC_FILLIO flag to
fetch the proc field "/proc/[pid]/io" data
process-wise.
IO Accounting is not included as part of procps.
Requirement:
We have a requirement to fetch process wise
IO utilization which can be used for monitoring.
When looking through the procps library, I see
that IO Accounting (/proc/[pid]/io) is not being
included as part of procps. There is no such
flag like PROC_FILLIO being included in readproc.h .
Solution:
While looking at the implementation done for
other proc fields, I used the spare bits in app code.
I renamed PROC_SPARE_1 as PROC_FILLIO, the spare bit from
PROC_SPARE_* and used it for fetching /proc/[pid]/io
data as part of the procps library similar to other
fields. I moved the PROC_SPARE_* bits each by 1 bit
to retain the spare bits. Meanwhile added the IO fields
in proc_t structure.
References:
procps-ng/procps!122procps-ng/procps#184
This simple two line code change fixes an intermittent
bug whereby %CPU for parent(s) with collapsed children
could be vastly understated from those displayed under
the current 3.3.17 publicly available top & libprocps.
If one started several top instances in the background
using very a small delay interval (zero?), then if the
shell under which they were running was collapsed, you
would see similar %CPU results for both the libraries.
However, when running a demanding 'make' like a kernel
compile (especially if backed by fast processors and a
SSD), then newlib would generally show only 1/3 to 1/2
of the collapsed %CPU values that appeared for 3.3.17.
Of course, now that the bug has been swatted with this
commit the disparities between those results is easily
explained. Since newly created tasks never contributed
tics during the interval where they were created, only
with many short lived tasks would differences surface.
Signed-off-by: Jim Warner <james.warner@comcast.net>
Rather than run risks of more expensive and repetitive
address resolution, we will establish this local index
for a one time cost and avoid any potential gcc bloat.
[ this commit was made in pursuit of a bug involving ]
[ the distortion of elapsed task tics. but, it turns ]
[ out these changes had nothing to do with that bug. ]
[ however, the patch is being retained as desirable. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
That old library defined this field as 'unsigned int'.
However, here it was known as a 'signed int'. Thus for
consistency we'll now also treat it as 'unsigned int'.
[ this commit was made in pursuit of a bug involving ]
[ the distortion of elapsed task tics. but, it turns ]
[ out these changes had nothing to do with that bug. ]
[ however, the patch is being retained as desirable. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
After removing brackets from those 'derived' notations
I was surprised to discover that several origin/source
comments were wrong. So this patch fixes those errors.
[ along the way a couple enumerators were renamed to ]
[ better (i hope) reflect what they're representing. ]
[ that, in turn, also required a little rearranging. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
Since 'procps_uptime' will access the /proc filesystem
the <pids> 'new' guy should should protect against the
possibility /proc isn't mounted when 'boot_seconds' is
established. A zero is better than the negative value.
[ the only distortion would be to PIDS_TIME_ELAPSED. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
In addition to copyright date, the initial descriptive
line was changed from a generic statement to one which
reflects the specific portion of the proc/ filesystem.
[ such descriptions alternate between 'declarations' ]
[ (h files) & corresponding 'definitions' (c files). ]
Also, a few missing copyright attributions were added.
Signed-off-by: Jim Warner <james.warner@comcast.net>
This commit just ensures that the relatively expensive
ID to name conversions aren't performed unless they're
explicitly requested. It also internalizes those flags
that required the PROC_FILLSTATUS flag to also be set.
[ requiring a caller, in our case pids.c, to provide ]
[ two flags when a single field was the objective is ]
[ wrong & represents a future potential toe-stubber. ]
[ moreover, what's worse is that those two flags are ]
[ seemingly unrelated. but, without both, a SEGV can ]
[ can be expected when a result.str pointer is NULL. ]
[ by contrast, in the master branch those fields are ]
[ arrays which, when set to zeroes, produce an empty ]
[ string. So, there is no abend (but no name either) ]
[ when one of those two required flags were omitted. ]
[ and worth noting, in that branch it's not just one ]
[ caller required to observe a two flag requirement. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
With the 4 header files removed in the previous patch,
this commit just changes all those obsolete references
to that new consolidated 'misc.h' header file instead.
Signed-off-by: Jim Warner <james.warner@comcast.net>
Under the above #define this commit now also addresses
2 additional possible toe stubbers involving 'select'.
If some readproc.h constants were uncoupled from their
pids.h enumerators a 'make check-lib' will now detect.
Signed-off-by: Jim Warner <james.warner@comcast.net>
If a hash results report is output (via UNREF_RPTHASH)
a portion is devoted to occupied table entries ordered
by depth. There is a possibility that some depths will
not be found among existing occupied table entries and
to avoid any confusion probably should not be printed.
[ to illustrate the potential for confusion prior to ]
[ this patch, force a very small table size (like 8) ]
[ & then trigger the procps_pids_unref() eoj report. ]
So this patch ensures only 'in use' entries are shown.
[ admittedly, all of the remaining logic in the loop ]
[ could/should be subordinate to this new 'if' test, ]
[ but we will keep the change to a minimum. besides, ]
[ there's no harm subtracting/adding a zero numdepth ]
[ especially since the chance of a zero is very low. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch just raises the size of the hash table used
to calculate elapsed task stuff. The net result should
be less need for 'chaining' under pid hash collisions.
[ the hash scheme is intentionally kept as primitive ]
[ and, therefore, as fast as possible. it employs an ]
[ 'and' approach versus a 'mod' operation since both ]
[ yield similar distribution but the former approach ]
[ was 4 fewer cpu instructions in terms of overhead. ]
[ additionally, for hash collisions, 'chaining' uses ]
[ an array index rather than the usual pointer since ]
[ the HST_t guys may move when they are reallocated. ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch separates the memory allocations into those
used initially from those used in later reallocations.
Thus, we can reduce that iterative realloc() overhead.
Additionally, we'll correct a long standing oops where
multiple history_info structures were created at 'new'
time when only one should have been allocated (jeeze).
[ originally the allocation was strangely based upon ]
[ number of 'items' (???) & later a #define constant ]
Reference(s):
. May, 2016 - subsequent bad history_info logic
commit 9ebadc1438
. Aug, 2015 - original faulty history_info code
commit 7e6a371d8a
Signed-off-by: Jim Warner <james.warner@comcast.net>
There's a huge toe-stubber awaiting future maintenance
as reflected in that commit below which deals with the
addition of new enumerators to the Item_table. Namely,
whenever the table is grown, one must remember to also
change that existing 'logical_end' enumerator's value.
Well, not anymore! Since that MAXTABLE macro was added
to the procps-private.h header we can now also exploit
it so a 'logical_end' automatically tracks table size.
This change also renders some code associated with the
ITEMTABLE_DEBUG #define unnecessary. So it's gone too.
Reference(s):
. 08/2016, add new enumerators
commit 09e1886c9e
. 08/2020, added MAXTABLE macro
commit c865b06c30
. 08/2020, introduced ITEMTABLE_DEBUG
commit 92d0297e1e
Signed-off-by: Jim Warner <james.warner@comcast.net>
The recent work on updating the <meminfo> and <vmstat>
modules with some newly added linux fields reminded me
(again) of a need for some mechanism guaranteeing that
a header file agrees with the source file assumptions.
Sadly, in the past, if a table entry was omitted or if
the table and header are ordered differently, then the
library would silently return the wrong results values
or even potentially experience a SIGSEGV abnormal end.
This patch offers a much needed development assist for
ensuring that Item_table entries are synchronized with
header file enumerators in terms of number plus order.
It's intended solely for our use as libprocps evolves.
Now, by activating ITEMTABLE_DEBUG, either directly or
via ./configure CFLAGS='-DITEMTABLE_DEBUG', the number
and order will be verified. It is envisioned that this
feature will be used at least once prior to a release.
Signed-off-by: Jim Warner <james.warner@comcast.net>
I almost feel like I should apologize for this kind of
patch ( *almost* ). But, since this comment was unique
to the <pids> api and since it was especially designed
to align with the following comment and since the next
comment already carried the alignment emphasis, I will
refrain from issuing any apology and submit it anyway.
[ plus to prove that i am not totally anal-retentive ]
[ and can exercise some restraint there are two more ]
[ places where i COULD have added alignment emphasis ]
[ see lines 804-806 and lines 1360-1361 for my proof ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch just trades the 'typestr' for the much more
meaningful 'valid range' when a relative enum has been
deemed invalid under that XTRA_PROCPS_DEBUG provision.
[ and we'll also make one comment a bit more generic ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch will just correct an oops introduced in the
commit shown below. Thank goodness both 'str' & 'strv'
occupy the same storage location in that result union.
Reference(s):
. standardize 'errno' management
commit 06be33b43e
Signed-off-by: Jim Warner <james.warner@comcast.net>
Some parts of our newlib implementation are the result
of functions which have been propagated from module to
module. In particular, those 'cleanup_stacks' routines
are all similar & likely originated in the <pids> api.
In that interface there was a need to free dynamically
acquired memory before the result structure was reused
to satisfy subsequent 'get', 'select' or 'reap' calls.
This, in turn, led to a concept of 'dirty' stacks with
the need to call one of two 'cleanup_stack' functions.
None of the remaining interfaces deal with such memory
yet they each had their own 'cleanup_stack' functions.
Those functions were responsible for resetting each of
the result unions to zero, excluding any 'noop' items.
The bottom line is that for all interfaces, repetitive
calls would require iterating through the stack(s) two
separate times: once to 'cleanup' another to 'assign'.
With this commit we will reduce iterations to just the
'assign' routine. A reset to zero will be accomplished
in the 'extra' item set routine (which is the only one
actually requiring any reset). All other items will be
reinitialized automatically by a new current set value
or upon reallocation when an items compliment changes.
In the <pids> interface, any freeing of dynamic memory
could have been accomplished by adding that 'freefunc'
check to the 'assign' function. However, that requires
an Item_table test with every item. Instead, we'll now
satisfy such needs as the very first step in those set
functions responsible for dynamically acquired memory.
[ the <pids> api retains 2 'cleanup_stack' functions ]
[ to accommodate stack(s) 'reset' & to serve 'unref' ]
Lastly, all the 'itemize_stack' functions were tweaked
by eliminating an unnecessary initialization of result
unions. That objective was already accomplished by the
calloc() in a 'stacks_alloc' function or the remaining
'cleanup_stack' routine found in the <pids> interface.
Signed-off-by: Jim Warner <james.warner@comcast.net>
When the <pids> api was refactored in the commit shown
below, one objective was enabling the simultaneous use
of 'get' & 'select/reap' functions. Unlike other 'get'
functions, this <pids> 'get' acts as an iterator where
successive calls will return successive tasks/threads.
However, that goal wasn't quite met since a stack used
by 'get' was commingled with the 'select/reap' stacks.
Such commingling supported the 'reset' function, again
a provision which was unique to this <pids> interface.
Unfortunately, some poor assumptions in 'stacks_fetch'
produced a SEGV whenever 'reap/select' followed 'get'.
Thus, this patch addresses those issues and guarantees
such commingled stacks (extents) will be accommodated.
Reference(s):
. standardize portions of interface, <PIDS> api
commit 9ebadc1438
Signed-off-by: Jim Warner <james.warner@comcast.net>
The commit referenced below addressed (some) anomalies
surrounding 'strv' pointers. However, there remained a
couple quirks involving a potential NULL return value.
Any NULL values returned from the old library readproc
guys would cause no real harm for newlib. But they did
produce the misleading "[ duplicate ENUM_ID ]" result.
The following all represent potential NULL results and
suggest shortcomings in testing of that earlier patch.
. kernel threads do not have cgroup, cmdline & environ
. even if present environ could require root to access
So, this patch reverts a portion of the earlier commit
and ensures when some vectored string is not available
a traditional dash ('-') is the 'strv' returned value.
[ and we'll also correct one typo in the header file ]
Reference(s):
. eliminated a final potential NULL, <PIDS> api
commit 09503dc597
Signed-off-by: Jim Warner <james.warner@comcast.net>
This commit is intended as a refinement of the patches
mentioned below, where origins/sources of newlib items
were added to the header files for user documentation.
However, if those additions are to be truly effective,
along with kernel documentation (where available), the
following prerequisites must also have been satisfied:
. our identifiers closely align with linux field names
. our derived items are documented or self-documenting
Satisfying those prerequisites prompted this patch and
for these changes, kernel sources were emphasized over
available documentation (shame on me, it should always
have been so). And, while some 'new' fields were found
to be conditional, they were included unconditionally.
These changes appear more extensive than they actually
need be since I have attempted to enforce some spacing
conventions. So, I've summarize the significant things
in the sections that follow. For a proper perspective,
use: 'git diff --ignore-space-change' (good as alias).
___________________________________________ <PIDS> api
This api is unique in that there exists many different
file/directory origins subordinate to /proc/<pid>. And
our item identifiers are sometimes coerced so as to be
able to group related or similar enumerators together.
So, users needed more help relating our identifiers to
an actual documented field. Thus, we will now also add
the field names as with 'stat: delayacct_blkio_ticks'.
Each item ending with a '_C' now consistently includes
both the parent's count/time plus waited for children.
That 'RTPRIO' guy was renamed/relocated as PRIORITY_RT
since its original name is an implementation artifact.
___________________________________________ <STAT> api
The only api change was to correct a typo ('dervied').
_________________________________________ <VMSTAT> api
Even ignoring white space, this interface received the
largest number of changes. Mostly, this was because of
deficiencies in the proc(5) documentation. Recall that
this documentation already sorely lacks any substance.
Usually, just kernel releases are noted, not contents.
When compared to kernel source, that proc(5) contained
many non-existent fields and also omitted many others.
________________________________________ <MEMINFO> api
Sadly, with this api many of the changes were simply a
correction of some earlier 'human error' where several
fields where hashed then tracked but never represented
with an item enumerator in this meminfo.h header file.
_______________________________________ <SLABINFO> api
The 'SLABS' (summary) & 'SLABNODE' items were reversed
since the former are derived from the separate caches.
More significantly, those 'SLABNODE' guys were renamed
to 'SLAB' since they concern individual caches and the
concept of 'nodes' is really an implementation detail.
Also, several enumerators were changed to more closely
agree with official slabinfo(5) documentation referred
to in what we're treating as a base document: proc(5).
Lastly, while those 'SLABS' items are solely a product
of our library and not represented in slabinfo(5), the
names attempt to parallel those found as 'SLAB' items.
______________________________________ <DISKSTATS> api
One enumeration identifier was changed so as to better
reflect its relationship to that actual documentation:
'Documentation/iostats.txt', as referenced in proc(5).
Reference(s):
. 12/2018, item origins added (and commit msg history)
commit 96d59cbf46
. 01/2019, <stat> origins tweaked
commit 201e816b26
Signed-off-by: Jim Warner <james.warner@comcast.net>
This change is being made in anticipation of adding the
source origin of each item to the <pids.h> header file.
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch is the first of three implementing a newlib
branch version of that Jan Rybar master merge request.
With this series we'll ultimately extend 'EXE' support
to both ps and top (plus, everyone else who wants it).
Reference(s):
. master branch merge request
https://gitlab.com/procps-ng/procps/merge_requests/66
Signed-off-by: Jim Warner <james.warner@comcast.net>
This detects an integer overflow of "strlen + 1", prevents an integer
overflow of "tot + adj + (2 * pSZ)", and avoids calling snprintf with a
string longer than INT_MAX. Truncate rather than fail, since the callers
do not expect a failure of this function.
---------------------------- adapted for newlib branch
. logic is now in pids.c
. former 'vectorize_this_str' is now 'pids_vectorize_this'
Signed-off-by: Jim Warner <james.warner@comcast.net>
With the documentation update in the commit referenced
below, we should also account for such threads as they
will already be represented in the task/thread totals.
[ and do it in a way that might avoid future changes ]
Reference(s):
commit 91df65b9e7
Signed-off-by: Jim Warner <james.warner@comcast.net>
This commit removes some obsolete parameter validation
code which was needed back when certain functions were
public, called directly by users (1st/2nd generation).
Now that they're static they can be safely eliminated.
Signed-off-by: Jim Warner <james.warner@comcast.net>
With older library logic having been modified to avoid
using those potentially deadly alloc.h routines, while
improving 'errno' handling, we're ready to standardize
and enhance newlib's approach to any potential errors.
In so doing, we'll establish the following objectives:
. . . . . . . . . . . . . functions returning an 'int'
. an error will be indicated by a negative number that
is always the inverse of some well known errno.h value
. . . . . . . . . . . functions returning an 'address'
. any error will be indicated by a NULL return pointer
with the actual reason found in the formal errno value
And, when errno is manipulated directly we will strive
to do so whenever possible within those routines which
have been declared with PROCPS_EXPORT. In other words,
in the user callable functions defined in source last.
[ But, that won't always be possible. In particular, ]
[ all the 'read_failed' functions will sometimes set ]
[ 'errno' so that they can serve callers returning a ]
[ NULL or an int without duplicating a lot of logic. ]
[ Also, that includes one subordinate function which ]
[ was called by 'read_failed' in the <slabinfo> API. ]
------------------------------------------------------
Along the way, several additional miscellaneous issues
were addressed. They're listed here now for posterity.
. the '-1' return value passed outside the library was
eliminated since it would erroneously equate to -EPERM
. the stacks_fetch functions in <diskstats> and <stat>
weren't checked for their possible minus return values
. hash create was not checked in <meminfo> or <vmstat>
. fixed 'new' function faulty parm check in <slabinfo>
Signed-off-by: Jim Warner <james.warner@comcast.net>
. ensure whitespace exists between the code & comments
[ changing txt slightly keeps right margin alignment ]
. strive for more consistency with some comment styles
[ don't use C '/*' style where C++ '//' style exists ]
. removed the instance of double space in 1 assignment
[ still striving for consistency in whitespace usage ]
. fixed comment relating to number of 'derived fields'
[ the <meminfo> api recently added one new such enum ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch just rearranges 1 item in the stacks_extent
struct to make it equivalent to all the other modules.
Signed-off-by: Jim Warner <james.warner@comcast.net>
In response to that suggestion referenced below, these
changes allow display of task/thread level NUMA nodes.
Currently, only the 'top' program offers any NUMA type
support and it is limited to the Summary Area display.
With this commit both the 'top' and 'ps' programs will
be able to display NUMA nodes associated with threads.
Reference(s):
https://gitlab.com/procps-ng/procps/issues/58
Signed-off-by: Jim Warner <james.warner@comcast.net>
In each module employing a priming read at 'new' time,
should that read fail, a call to 'unref' will be made.
However, there is a hidden dependency that these calls
must never occur before the context 'refcount' was set
due to the way an 'unref' conditional was constructed.
So this commit just ensures that 'unref' will function
as expected, even if called with a 'refcount' of zero.
Signed-off-by: Jim Warner <james.warner@comcast.net>
Removing the Item_table 'stat' oldflags for WCHAN_ADDR
was wrong since that 'stat' field is not a constant 0.
Rather, it could assume these 3 values: -1, 0, and +1.
I have not been able to pin down a '-1' result, but it
probably means some sort of permission error (-EPERM).
The '1' or '0' values were supposed to distinguish the
tasks that were or were not blocked (whether there was
a wchan address). However, in practice there is little
correlation between those values and availability of a
kernel symbol in /proc/$$/wchan (perhaps due to race).
Anyway, the real point is that a 'stat' wchan does not
now intentionally contain an address. Thus, outputting
'ffffff', '-' or '1' in programs like ps is senseless.
So this patch just eliminates PIDS_WCHAN_ADDR from our
item enumerators leaving only the PIDS_WCHAN_NAME guy.
Now the new library can't be blamed for bad addresses!
Reference(s):
. removed Item_table 'oldflags'
commit c4aa6c0ab4
. linux removal of wchan addresses
commit b2f73922d119686323f14fbbe46587f863852328
Signed-off-by: Jim Warner <james.warner@comcast.net>
It seems inappropriate to blindly include fields known
to always be zero in our brand new library. Therefore,
this patch removes support for three such enumerators.
[ that stat 'it_real_value' (PIDS_ALARM) field could ]
[ have been made obsolete before a linux 2.6 release ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This commit just contains some tweaks to comments plus
a few adjustments to whitespace for alignment purposes
and a normalization of the header inclusion #define's.
[ plus a spelling error in one header file was fixed ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
Recent profiling and timings have resulted in improved
newlib performance. This patch completes that process.
It just normalizes naming conventions employed for all
allocations involving reaped stacks & history support.
The modules offering a 'reap' function will also offer
the now standardized corresponding STACKS_INCR define.
The modules which provide dynamic history support will
now have a separate #define called NEWOLD_INCR used in
allocations/reallocations. And, while values currently
are set equal to that STACKS_INCR value, in the future
some reason for divorcing those two may be discovered.
----------------------------- for future reference ---
In those modules which contain the STACKS_INCR #define
it is tempting to specify a large value so as to avoid
repeated calls to malloc/realloc. However, in doing so
an extra runtime price will be paid in 'cleanup_stack'
calls with any iterative programs like top or slabtop.
So, with the current values a balance has been sought.
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch contains the following miscellaneous stuff:
. The pids_stacks_fetch() routine might call for newly
allocated stacks to be itemized. However, that job was
already tended to by the pids_stacks_alloc() function.
So, this patch just eliminates a redundant invocation.
------------------------------------------------------
. The concept of 'dirty_stacks' has not kept pace with
the evolving stacks implementation. Originally, stacks
were considered dirty only if free() of dynamic memory
was needed before refreshing any single result struct.
Later, with the introduction of the 'extra' item and a
promise to reset it to zero, 'dirty' was much broader.
So, this patch just treats the dirty flg as others do.
------------------------------------------------------
. Lastly, a word or three about performance & timings.
Tuning efforts concentrated on the <pids> API and top.
And unless an oldlib equivalent to the preceding patch
is applied (favoring stat vs. status), newlib top will
often outperform the oldlib version (obviously wrong).
So assuming /proc/stat is preferred in both libraries,
generally speaking, a cpu and elapsed time increase of
1-5% was found for this new stacks oriented interface.
Of course, there's also an increased memory footprint.
There are some occasions, however, when the newlib top
is at a substantial disadvantage. For example if WCHAN
or TTY is displayed, such items will be present in all
newlib reaped stacks (i.e. every process). But old top
would only incur such overhead with displayable tasks.
Thus, oldlib top could outperform newlib by up to 25%,
for example, if only fields requiring NO library flags
were displayed. However, such a scenario is not likely
since only GID, UID, PID, TGID & WCHAN would be shown.
In the usual case, that overhead associated with WCHAN
and/or TTY is overshadowed by other top runtime costs.
All in all a pleasing outcome I deem quite acceptable.
------------------------------------------------------
Signed-off-by: Jim Warner <james.warner@comcast.net>
Long ago, in a galaxy far away (when top was in charge
of library FILL flgs) /proc/status was to be preferred
over /proc/stat if a field could be satisfied by both.
This was done to avoid costly 64-bit math emulation in
a 32-bit application due to 'unsigned long long' data.
Well it's time to acknowledge the prevalence of 64-bit
platforms. And in such an environment the cost picture
has shifted significantly. It now costs 14 times (wow)
as much to access /proc/status compared to /proc/stat.
In other words, even with '%llu' variables, a sscanf()
call in stat2proc() beats the pants off that home brew
gperf based hashing employed by the status2proc() guy.
In fact, status2proc incurs higher costs than found in
the most expensive aspect of top's forest view option.
Here's a gprof extract to illustrate the costs. It was
produced with an rcfile requiring fields from both the
/proc/stat & /proc/status pseudo files (among others).
There were 5000 iterations in each of 4 separate gprof
runs subsequently merged into 1 gmon.sum for analysis.
% self self
time seconds calls us/call name
----- ------- ------- ------- -----------
28.65 4.10 4689423 0.87 status2proc
26.14 3.74 40000 93.50 forest_adds
...
01.96 0.28 4689427 0.06 stat2proc
[ since forest_adds is recursive, the calls value is ]
[ the non-recursive #, its 'call graph' shows totals ]
Anyway, now that such cost is known this patch becomes
what is euphemistically known as the usual no-brainer.
[ jeeze, was it really this long between profilings? ]
Signed-off-by: Jim Warner <james.warner@comcast.net>
This patch will begin some refinements associated with
gprof. Initially, functions names have been changed to
help in identifying potential bottlenecks. This effort
also included the obscure set, free and sort routines.
Plus the following additional modifications were made:
. the stacks_alloc prologue was generalized plus added
to a couple of modules where it had not yet propagated
. a couple of the '// end ...' comments were corrected
. some functions have been formally tagged as 'inline'
Signed-off-by: Jim Warner <james.warner@comcast.net>
Rather than return a 0 result for all VAL type errors,
return what would have been without validation active.
This will enable a program like pgrep to still print a
result even though it used some incorrect type member.
With this commit, our VAL macro validations logic will
behave in exactly the same way as the GET validations.
While warning messages may be issued, except for a bad
enumerator, values will always be returned to callers.
Signed-off-by: Jim Warner <james.warner@comcast.net>
