vmstat.c:
Move totfr to be under daefr and prcfr since it logically belongs there.
Move all the count fields (wire, act, inact, cache and free) to near
the bottom of the sub-display (after all the rate fields) to reduce
competition with adjoining sub-displays.
systat.1:
Move things as above.
Attempt to improve missing and poor wording in the description of the
fields. The long sentence was hard to parse and didn't say anything
about the different units.
Increment .Dd.
part that handled the 17th and 18th rows of the vmstat-proper subdisplay
was deleted in rev.1.10 when these rows stopped being used and was not
restored when the 17th row was used again. For such terminals, we now
lose the `buf' field instead of making a mess with it. Terminals with
fewer than 24 rows have never been supported.
The problem is not avoided by using curses since we use the last line
for data entry and don't use a separate subwindow for this line.
Some other things in the vmstat display could be handled better using
subwindows.
output too.
Fine tune all coordinates and most field widths in the vmstat (sub)display
for this and previous changes now that we have to change almost all of them
just to move the ex-extended fields:
- change VMSTATROW back to 7. It was 6 due to a hack in the extended vm
stats changes.
- reduce the maximum field width that we try for from 9 to 8. 4 or 5 is
enough for most fields but we try to use the same width for all fields.
8 is enough to display everything without changing units memory sizes
exceed 100GB.
Fix some unrelated coordinates and field widths in comments.
vm stats to the normal vm stats. Sort them into the normal stats
according to the man page only in the source code so that diffs are
almost readable. Reduce style bugs in printing the value of %ozfod.
new vnstat display to the right of the namei display.
Move the non-vmstat fields {des,num,fre}vn from the vmstat display to a
new vnstat display. Move the dtbuf field there too. The buf and dtbuf
fields are non-vmstat and non-vnstat, so there is no good place to
display them. I need to move at least 1 of them out of the vm stats
for further cleanups of the vm stats, and there is only space for 1
of them in the vn stats. (The best place for the current buf field
is actually /dev/null, since it has been completely broken for about
10 years and broken for longer. It gives an uninteresting virtual
memory count where an interesting real memory count is wanted.)
to handle changes to the set of disks selected, but it is unnecessary
for that since the whole screen is redrawn when this set is changed.
It was also buggy:
- MAXDRIVES*6 = 42 was hard-coded as only 30 spaces in a string literal,
the last 2 disk names were not cleared as intended
- when the extended vmstats are active, clearing of even 30 columns
overruns the ozfod value field by 3 columns. This was harmless because
the field is much wider than necessary.
value printed is actually the optimized (i.e., the non-slow, not-on-the-fly
zero fills percentage) except in overflow cases. Describe it as %ozfod
in the display. Move the field descriptor 1 to the left so that there
is space for 5 characters after the % sign (this leaves no space between
the number and the descriptor but the % character serves well as a
separator).
Fixed integer overflow at z.ozfod = UINT_MAX/100 in the calculation of
%ozfod. This value can be reached just a few hours or minutes after
booting, so %ozfod was usually garbage in boot mode. Now %ozfod is
correct in boot mode for a few days or hours.
Print a non-dummy %ozfod when the division for it isn't division by 0
instead of when the result will be less than 100%. A result of 100%
may be correct, though a result of more than 100% indicates overflow
of one or both counters.
not very usefully, in all other displays). This was the original point
of the PR.
Move the load average up by 2 so that it starts in row 0 for all windows
(2 lines above it were wasted for all other windows except vmstat).
Move everything below it up by 2 or 3 (3 for icmp and icmp6 which had
an extra blank line due from not compensating for the foot-shooting in
note (3); only ip and ip6 compensated). Reduce the magic numbers related
to this.
Notes by the submitter:
%%%
1. All the subwin() calls are identical using #define MAINWIN_ROW 3
(systat.h).
2. The load average is at the top of the window.
3. Each display starts on the fourth line. I made changes to those
displays that shifted the start line (i.e., icmp). This entailed a
lot of changes within the comments at the top of those displays.
4. For ip6, I shifted the "Input next-header histogram" column down one
row to separate it from "IPv6 Output". I raised "bad scope packets"
and "address selection failed" up one row to stay with "IPv6 Input"
(valid?). They were down one row to probably line up at the bottom,
but I think they should stick with their fellow items in a column.
5. I condensed ifstat a bit. It had a lot of empty rows.
%%%
Submitted by: Se=E1n Farley <sean-freebsd at farley dot org>
PR: bin/81874
only affect amd64 and i386. alpha uses "intr N" instead of "irqN" and
mostly has no device names. ia64 uses only device names.
- Edit interrupt names once after they are read from the kernel and not
every time they are displayed.
- Discard bogus trailing spaces so that the next step doesn't move things
to oblivion.
- If an interrupt name starts with "irqN:" (as it usually does in on
amd64 and i386), then move "irqN" to the end and strip ":", since we
have no space for the ":" and don't want to start descriptions with
"N" after stripping "irq" in the next step (since "N" would look like
a count). This step may need reworking for interrupt names containing
several device names -- then moving the irq number to the end would
lose it instead of losing some device names.
- Remove "irq" from an interrupt name if and only if the original name is
too long to display.
accidentally.
Read buffer overruns:
The size of the target array (TSOTTA == 10) is a wrong limit to use for
scanning the source string.
Write buffer overruns:
TSOTTA is also a wrong limit to use for copying to the target buffer,
since we want to add a NUL terminator afterwards. TSOTTA was also 1
too small for holding both the desired number of visible characters
and the NUL.
Worked accidentally:
There is error in the algorithm that tends to result in the space saved
by stripping "irq" not actually being used, but some cases worked
accidentally provided "irqN" is near the end of the source string and
"N" is only 1 digit.
Starting with 5.mumble-CURRENT, "irqN" is at the beginning of the
string on all (?) arches that have it and the accidents don't happen.
E.g. on i386's, the keyboard irq is now named
"irq1: atkbd0<bogus blank padding>" by the kernel, and this name was
converted to "1: atkb" -- not only the device number but part of the
device name has been lost --, while before 5.mumble the kernel name
was "atkbd0 irq1" and systat accidentally preserved the irq number to
give "atkbd0 1". The ":" in the string wastes precious space, and
stripping "irq" results in descriptions starting with numbers which
makes them look too much like counts. This commit just fixes the last
problem.
clobbered at runtime:
dirtybuf -> dtbuf
desiredvnodes -> desvn
numvnodes -> numvn
freevnodes -> frevn
The vmstats column has only 5 characters available for descriptors, but up
to 13 were used. The extras get clobbered at runtime by interrupt values
and/or descriptors on systems with more than 12 interrupt sources.
%slo-z -> %sloz
This one is in the "extended" vmstats area and doesn't get clobbered now.
Removed stale documentation of desvn.
Changed a descriptor:
tfree -> totfr
so that it is consistent with the abbreviations for other free counts
(daefr and prcfr) and thus almost decodeable.
Fixed missing documentation of tfree/totfr. This and everything else
in the extended vmstats area is misdocumented as being in a certain
place in the vmstats column.
w/ non-zero data, and it turns out we don't... This is really optimized
zero filled on demand, or pages that were already zero'd for us...
MFC after: 3 days
3 columns were wasted at the left, except these columns were used to
make the header line up. Now there is no space on the same line for
the "Proc:" part of the header. Try putting this on the line above
although it clutters that line (there is already similar clutter for
the "Interrupts" header). Leave 1 column between these fields. With
the above and a previous change there is enough of space for this.
Use 5 columns instead of 3 for the number of users since 3 is not quite
enough and there was space to spare. This also fixes an off-by-2 error
in a previous fix forthe column count in the comment on STATROW.
Move all the pager fields 1 to the right so that the "count" and "pages"
descriptors more clearly apply to the pager fields and not the memory
fields. There was space to space.
Waste some of the spare space at the right of the pager fields to expand
all the pager field widths to their old values (but now with a column
between the fields). There are fields more in need of expansion but most
of them are not in places near spare space.
made it unnecessary. (Rev.1.6 had to reduce the field width to 4, and
changed 100.0 and preposterous larger values down to 99.9 since 100.0
wouldn't have fitted. Rev.1.35 handles precentages > 99.9 well enough by
changing the format to %.0f when the string given by the initial format
is too wide.)
Even with this change, during short testing I've never seen a percentage
of 100 being displayed by systat -v, although top(1) displays percentages
of 100 user or 100 idle for similar loads.
Always use snprintf()'s return value, since discarding it is a style
bug at best and using it here gives slightly simpler code and better
error checking. Use snprintf() in putlongdouble() the same as in
putfloat(). (1.25 changed most sprintf()'s to snprintf()'s to fix
non-bugs without changing the logic to use the result of snprintf();
1.27 restored one of the sprintf()s by cloning a stale version of
putfloat().)
Don't print a too-long field in the unlikely case that the fallback
to M units in putint() leaves the field still too long. (The fallback
to printing stars was lost in rev.1.58 when the fallback to M units
was added.)
cannot run into other fields or field descriptors. If the value is
too large to fit in the field width, then the output format is adjusted
so that the value (usually) fits, but with fields running together
externally this adjustment usually didn't help. Mostly it doesn't
matter to lose 1 digit of precision, but switching the output format
is bad if it happens often or gives bogus units. The loss of width
is most serious for fields near "Csw" (which are also the ones which
must often ran together) since these have a high variance and large
values relative to the possible field widths so the switch occurs more
often now, and for the memory size fields where the switch gives the
bogus units kKB or MKB.
Now only the fields for r, p, d, s and w can run into each other.
These fields have width 3, and 3 cannot be reduced to 2 without losing
all precision when the value is between 100 and 999.
Trim "pdwake" to "pdwak" at think time now that it doesn't get clobbered
at runtime. The manpage doesn't need to be changed for this because
it documents the clobbered descriptor, unlike for 4 other too-long
descriptors which only get clobbered if there are lots of interrupt
sources.
Trim "% busy" to "%busy" since most other descriptors for percentages
are spelled without the space and this change makes changing the widths
of the %busy fields unnecessary.
around PUTRATE() because PUTRATE() only looked like a function -- it was
multiple statements. Use "do {...} while(0)" as usual in PUTRATE() so
that it is a single statement that can be used like a function.
large. In most cases it is still 1 too large, so fields tend to run
together, but in the following cases it was more than 1 too large, and
the starting column was too small too, so the field started inside the
previous field or descriptor and clobbered that:
- "wire": the number for this overwrote 2 characters of the number for
"Flt". Reduce the field width by 3 (2 to avoid the overwrite and 1
so that the fields don't run together). This was already done for
the preceding number for "cow".
- "inact": the number for this overwrote 1 character of the descriptor
"Idle". Reducing the field width by 2 is enough.
- "cache:" the number for this overwrote 3 characters of the scale
"...| |". The field width should be reduced by 4 to keep things
from running together, but that is a lot and not so necessary here
since the final "|" in the scale serves as a delimiter. Only reduce
it by 3.
- "free": the number for this overwrote 2 characters of the bar graph.
The character position under the final "|" in the scale is apparently
not used, so reducing the field width by 3 is enough.
When "zfod" is in the main vmstat display:
- use the normal field width of 9 (not 5) for it since there is no shortage
of space. Fix style bugs (excessive {}) in the statement that
conditionally writes it.
Write all reduced field widths for vmstat fields as "9 - <reduction>" as
a hint that we don't want to reduce them.
number in more cases by stealing 2 characters from the count field to
give more space in the descriptor field, but it did the column adjustments
for this strangely using an off-by-2 error in the base column and
compensating off-by-2 errors in 6 offsets from the base column (4 new
errors and 2 from not changing the offsets that actually changed).
Print the "Interrupts" header directly at its offset from the base column
instead of spacing it half using the offset and half by printing a space
character.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
<netinet/tcp_var.h>'s prerequisites. Prerequistes should not grow for
userland headers, and <netinet/tcp_var.h> is unfortunately still needed
in userland.
where MB/s and tps statistics would always be zero, presumably because
they were being averaged out over the time between now and when the
system booted instead of a few seconds.
PR: 58683
Kernel:
Change statistics to use the *uptime() timescale (ie: relative to
boottime) rather than the UTC aligned timescale. This makes the
device statistics code oblivious to clock steps.
Change timestamps to bintime format, they are cheaper.
Remove the "busy_count", and replace it with two counter fields:
"start_count" and "end_count", which are updated in the down and
up paths respectively. This removes the locking constraint on
devstat.
Add a timestamp argument to devstat_start_transaction(), this will
normally be a timestamp set by the *_bio() function in bp->bio_t0.
Use this field to calculate duration of I/O operations.
Add two timestamp arguments to devstat_end_transaction(), one is
the current time, a NULL pointer means "take timestamp yourself",
the other is the timestamp of when this transaction started (see
above).
Change calculation of busy_time to operate on "the salami principle":
Only when we are idle, which we can determine by the start+end
counts being identical, do we update the "busy_from" field in the
down path. In the up path we accumulate the timeslice in busy_time
and update busy_from.
Change the byte_* and num_* fields into two arrays: bytes[] and
operations[].
Userland:
Change the misleading "busy_time" name to be called "snap_time" and
make the time long double since that is what most users need anyway,
fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same
timescale as the kernel fields.
Change devstat_compute_etime() to operate on struct bintime.
Remove the version 2 legacy interface: the change to bintime makes
compatibility far too expensive.
Fix a bug in systat's "vm" page where boot relative busy times would
be bogus.
Bump __FreeBSD_version to 500107
Review & Collaboration by: ken
