IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
If the group_id passed to _stats_group_id_present is equal to the
special value DM_STATS_GROUP_NOT_PRESENT there is no need to perform
any further tests: return false immediately.
Call _stats_regions_destroy() from dm_stats_list() if dms->regions
is non-NULL. This avoids leaking any pool allocations and ensures
the handle is in a known state: if an error occurs during the list,
dms->regions will be NULL and the handle will appear empty.
If FIEMAP returns a single extent after the first call, no extent
boundary is detected and the first extent is not counted by the
normal mechanism.
In this case, increment nr_extents at the same time the extent is
added to the region table, before returning.
It's useful to be able to specify a minimum number of bits for a
new bitmap parsed from a list, for e.g. to allow for expansing a
group without needing to copy/reallocate the bitmap.
Add a backwards compatible symbol for programs linked against old
versions of the library.
Split out the loop that iterates over each batch of FIEMAP
extent data from the function that sets up and calls the ioctl
to reduce nesting and simplify local variable use:
_stats_get_extents_for_file()
-> _stats_map_extents()
The _stats_map_extents() function is responsible for detecting
eof and extent boundaries and adding whole, allocated extents
to the file extent table for region creation.
Check that all region_id values specified in a group bitmap are
actually present: although this should not normally happen when
using the dmstats tool, it is possible as a result of manual
changes (or bugs) for a group descriptor to contain one or more
group_id values that do not exist.
Check for this situation when reading group descriptors, warn
the user the user, and clear these bits in the bitmap when
formatting it for output.
If a region has a a DMS_GROUP tag in aux_data where the first
region_id in the bitmap is not the same as the containing region,
dmstats will segfault:
# '2' is never a valid group bitset list for region_id == 0
# dmsetup message vg_hex/root 0 "@stats_set_aux 0 DMS_GROUP=img:2#"
# dmsetup message vg_hex/root 0 "@stats_list"
0: 45383680+16384 16384 dmstats DMS_GROUP=img:2#
1: 46071808+32768 32768 dmstats -
2: 47382528+16384 16384 dmstats -
# dmstats list
Segmentation fault (core dumped)
The crash will occur in some arbitrary dm_stats_get_* property
method - this happens while processing the 1st region_id in the
bitset, because the region is marked as grouped, but there is
no group bitmap present at dms->groups[2]->regions.
Fix this by detecting a mismatch between the expected region_id
and dm_bit_get_first() for the parsed bitset during
_parse_aux_data_group().
Handle files that contain multiple logical extents in a single
physical extent properly:
- In FIEMAP terms a logical extent is a contiguous range of
sectors in the file's address space.
- One or more physically adjacent logical extents comprise a
physical extent: these are the disk areas that will be mapped
to regions.
- An extent boundary occurs when the start sector of extent
n+1 is not equal to (n.start + n.length).
This requires that we accumulate the length values of extents
returned by FIEMAP until a discontinuity is found (since each
struct fiemap_extent returned by FIEMAP only represents a single
logical extent, which may be contiguous with other logical
extents on-disk).
This avoids creating large numbers of regions for physically
adjacent (logical) extents and fixes the earlier behaviour which
would only map the first logical extent of the physical extent,
leaving gaps in the region table for these files.
When mapping regions to a file descriptor, a temporary table of
extent descriptors is built using the dm_pool object building
interface.
Previously this use borrowed the dms->mem region and counter
table pool (since nothing can interleave with the allocation
while the caller is still in dm_stats_create_regions_from_fd()).
This turns out to be problematic for error recovery. When a
region creation operation fails partway through file mapping,
we need to roll back the set of already created regions and
this requires a listed handle: the dm_stats_list() will then
allocate from the same pool as the extents; we either have
to throw away valid list data, or leak the extent table, to
return the handle in a valid state.
Avoid this problem by creating a new, temporary mem pool in
_stats_create_file_regions() to hold the extent data, and
discarding it on exit from the function.
While cleaning up the table of already created regions during a
failed dm_stats_create_regions_from_fd(), list the handle once,
and call _stats_delete_region() directly. This avoids sending a
@stats_list message for each region deleted, reducing runtime
from 6s to 0.7s when cleaning up ~250 out of ~10000 regions:
# time dmstats create --filemap b.img
device-mapper: message ioctl on (253:0) failed: Cannot allocate memory
Failed to create region 246 of 309 at 9388032.
Could not create regions from file /root/b.img
<< pauses here >>
Command failed
real 0m6.267s
user 0m3.770s
sys 0m2.487s
# time dmstats create --filemap b.img
device-mapper: message ioctl on (253:0) failed: Cannot allocate memory
Failed to create region 246 of 309 at 9388032.
Could not create regions from file /root/b.img
Command failed
real 0m0.716s
user 0m0.034s
sys 0m0.581s
Testing the error path requires region creation to start to
fail part way through the operation (in order to have regions
to clean up): the simplest way is to ensure the system is
close to the kernel limit of 1/4 RAM or 1/2 vmalloc space
consumed by dmstats data.
Split dm_stats_delete_region() so that internal callers can manage
the handle state themselves.
dm_stats_delete_region() now just handles checking the state of the
handle, reporting validation errors, and calling dm_stats_list() if
necessary, before calling _stats_delete_region().
The new _stats_delete_region() function performs the actual group
member removal and region deletion, and requires a fully listed
handle to operate.
Callers that repeatedly delete regions can use a single listed
handle for many operations on the same device, avoiding one
message ioctl per region deleted: since @stats_list with many
regions is expensive, this yields large runtime improvements.
If we fail to create a region during dm_stats_create_regions_from_fd(),
we must remove all regions that were created to do this to date. This
needs to loop over the table of region_id values that were populated
by _stats_create_file_regions() before the error.
The code for this failure case in the out_remove branch incorrectly
uses the table index as the region_id:
for (--i; i != DM_STATS_REGION_NOT_PRESENT; i--) {
if (!dm_stats_delete_region(dms, i))
log_error("Could not delete region " FMTu64 ".", i);
}
This causes the cleanup code to delete a completely unrelated set
of regions (since the index here will always be nr_regions..0).
Fix it to pass the actual region_id stored in regions[i] instead.
Fix a silly bug in dm_stats_delete_region() that hugely inflates
runtimes when deleting a large number of regions.
For ~50,000 regions this change reduces the runtime from 98s to
6s on my test systems (a ~93% reduction).
The bug exists because dm_stats_delete_region() applies a truth
test to the return value of dm_stats_get_nr_areas(); this is
never correct usage - it will walk the entire region table and
calculate area counts for each region (which is roughly O(n^2)
in the number of regions, as dm_stats_delete_region() is being
called inside a region walk).
Although the individual area calculation is not that costly,
uselessly running anything 2,500,000,000 times over gets a bit
slow.
A much cheaper test (which is always true if the areas check is
true) is to just test dm_stats_get_nr_regions() or dms->regions;
if either is true it implies at least one area exists.
Old:
Performance counter stats for 'dmstats delete --allregions --alldevices':
98117.791458 task-clock (msec) # 1.000 CPUs utilized
127 context-switches # 0.001 K/sec
3 cpu-migrations # 0.000 K/sec
6,631 page-faults # 0.068 K/sec
307,711,724,562 cycles # 3.136 GHz
544,762,959,577 instructions # 1.77 insn per cycle
84,287,824,115 branches # 859.047 M/sec
2,538,875 branch-misses # 0.00% of all branches
98.119578733 seconds time elapsed
New:
Performance counter stats for 'dmstats delete --allregions --alldevices':
6427.251074 task-clock (msec) # 1.000 CPUs utilized
6 context-switches # 0.001 K/sec
0 cpu-migrations # 0.000 K/sec
6,634 page-faults # 0.001 M/sec
21,613,018,724 cycles # 3.363 GHz
3,794,755,445 instructions # 0.18 insn per cycle
852,974,026 branches # 132.712 M/sec
808,625 branch-misses # 0.09% of all branches
6.428953647 seconds time elapsed
There are two possible errors in _dm_stats_populate_region():
* No region struct in dms->regions[region_id]
* Failure to parse data from @stats_print
These have very different causes: the first occurs where a client
program is populating one region at a time (region_id is a single
region identifier), and has not previously called dm_stats_list()
to dimension the region tables; this is an API usage error.
The second occurs when either we read unparseable data from the
kernel (kernel bug), or where various resource allocations fail.
Separate these two cases out and log separate messages for each
(allocation failures in the path already have their own distinct
message), since the "failed to parse.." message in the un-listed
handle case is confusing and misleading.
Translate log_info() into log_very_verbose() which is macro
supposed to be used by our code.
log_info() is internal macro with eventually some 'symbolic' meaning
in syslogging daemons.
The dm_stats_delete_region() call removes a region from the bound
device, and, if the region is grouped, from the group leader
group descriptor stored in aux_data.
To do this requires a listed handle: previous versions of the
library do not since no dependencies exist between regions without
grouping.
This leads to strange behaviour when a command built against an old
version of the library is used with one supporting groups. Deleting
a region with dmstats succeeds, but logs errors:
# dmstats list
Name RgID RgSta RgSiz #Areas ArSize ProgID
vg_hex-root 0 0 1.00g 1 1.00g dmstats
vg_hex-root 1 1.00g 1.00g 1 1.00g dmstats
vg_hex-root 2 2.00g 1.00g 1 1.00g dmstats
# dmstats delete --regionid 2 vg_hex/root
Region ID 2 does not exist
Could not delete statistics region.
Command failed
# dmstats list
Name RgID RgSta RgSiz #Areas ArSize ProgID
vg_hex-root 0 0 1.00g 1 1.00g dmstats
vg_hex-root 1 1.00g 1.00g 1 1.00g dmstats
This happens because the call to dm_stats_delete_region() is inside
a dm_stats_walk_*() iterator: upon entry to the call, the iterator
is at its end conditions and about to terminate. Due to the call to
dm_stats_list() inside the function, it returns with an iterator at
the beginning of a walk and performs a further iteration before
exiting. This final loop makes a further attempt to delete the
(already deleted) region, leading to the confusing error messages.
The current dmsetup.c handles DR_STATS and DR_STATS_META reports
separately in _display_info_cols(), meaning that the stats walk
functions are never called for these report types.
Versions before v2.02.159 have a loop using dm_stats_walk_do() and
dm_stats_walk_while(), that executes once for non-stats reports,
and once per region, or area, for DR_STATS/DR_STATS_META reports.
This older behaviour relies on the documented behaviour that the
walk functions will accept a NULL pointer as the struct dm_stats*
argument.
This was broken by commit f1f2df7b: the NULL test on dms and
dms->regions were incorrectly moved from the dm_stats_walk_end()
wrapper to the internal '_stats_walk_end()' helper.
Since the pointer is dereferenced in between these points, using
an older dmsetup with current libdm results in a segfault when
running a non-stats report:
# dmsetup info -c vg00/lvol0
Segmentation fault (core dumped)
Restore the NULL checks to the wrapper function as intended.
Support aggregate group and region histograms by allocating a new
histogram from the pool and populating it with a sum of the histogram
data for the areas contained in the region or group.
To avoid repeatedly summing the same histogram data, cache the pointer
in the group and regions structs for subsequent access. The aggregate
histograms are allocated from the same pool as the area histograms in
the corresponding handle and will be discarded at each list or populate
operation.
Add a call to create dmstats regions that correspond to the extents
present in a file descriptor open on a file in a local file system.
The file must reside on a file system type that correctly supports
physical extent location data in the FIEMAP ioctl.
Regions are optionally placed into a group with a user-defined alias.
File systems that do not support physical offsets in FIEMAP (btrfs
currently) are detected via fstatfs() - although attempting to map
a --filemap group on btrfs will fail anyway with the generic error
"Not on a device-mapper device" this is confusing; the file system
mount is on a device-mapper device, but btrfs' volume layer masks
this in the returned st_dev field since the returned logical file
extents may span multiple physical devices.
The function _stats_remove_region_id_from_group() incorecctly set
the group_id to DM_STATS_GROUP_NOT_PRESENT _before_ the call to
_stats_group_destroy(). This will cause the destroy function to
return immediately without doing anything:
339 static void _stats_group_destroy(struct dm_stats_group *group)
340 {
341 if (!_stats_group_present(group))
342 return;
Invalidating the ID in _stats_region_region_id_from_group() is
redundant anyway; it is rightly done as the last operation in
_stats_group_destroy (and it is not possible for anything to see
the old value between the two calls).
Remove the change to group_id to ensure that the alias and bitset
resources are correctly freed.
If after extracting stats arguments and group tags nothing remains
of aux_data but '-' set the region->aux_data field to the empty
string to match behaviour for non-grouped regions.
Although not harmful do not allow a group containing regions with
histograms since it is not currently possible to present histogram
data aggregated for the group.
Although a non-zero value for the number of ticks spent doing IO
should imply a non-zero number of IOs in the interval test for
this explicitly to avoid a divide-by-zero in the event of bad
counter data.
It's possible for interval_ns to be zero if the interval is not
set or the clock is misconfigured. Test for this before using the
value as the divisor in the utilisation calculation.
Walk flags are ULL constants; cast the result to a uint64_t before
logging with a FMTx64 format specifier to avoid a compiler warning:
warning: format ‘%lx’ expects argument of type ‘long unsigned int’,
but argument 5 has type ‘long long unsigned int’
Make it clear in libdevmapper.h, and in function argument names, that
libdm-stats uses the aux_data field internally and that any values set
for user_data are appended to the library values before being stored
with a region, and similarly, that internal data fields will be stripped
prior to returning any previously stored user_data.
Add support do dm_stats_walk*() to walk over the set of
available groups using the cursor embedded in the dm_stats
handle, and to obtain the type of the object at the current
stats cursor location. A set of flags is introduced to
control which objects are visited:
DM_STATS_WALK_AREA
DM_STATS_WALK_REGION
DM_STATS_WALK_GROUP
DM_STATS_WALK_ALL
A final flag suppresses visits to regions that contain only a
single area - since the aggregate of such a region is idential
to the area it contains this allows these duplicates to be
filtered out:
DM_STATS_WALK_SKIP_SINGLE_AREA
If flags are not initialised before beginning a walk the default
set matches the behaviour of previous versions of the library.
Also accept group identifiers as immediate arguments to the
counter, metric, and property functions by adding control
flags to the region and area identifiers passed in.
Region and area properties are mapped to their equivalents for
the group (for example: group size is reported as the sum of
all regions contained in the group). Counter and metric values
are aggregated for the region or group.
Introduce constants for the buffer sizes that libdm-stats uses:
one for messages sent to the kernel, one for rows of response data
returned, and a pair for the "start+len" range and histogram bounds
strings.
Add a grouping facility to the libdm-stats library that allows the
user to bind several regions together as a group. Groups may be
used to aggregate data from several regions for reporting, or to
select and sort among large sets of regions.
A textual descriptor ("group tag") is associated with each group
and is stored in the first group member's aux_data field. The
tag contains the group member list and an optional alias for the
group, allowing the user to assign meaningful names to groups of
regions.
These descriptors are parsed in @stats_list message responses and
populate the resulting region and area tables with the group
structure.
Groups with overlapping regions are permitted but since this will
result in some events being counted more than once a warning is
printed in this case.
Nested and overlapping groups are not currently supported and
attempting to create these configurations results in error.
Add a new enum based interface for accessing counter and metric
values that uses a single function for each:
uint64_t dm_stats_get_counter(const struct dm_stats *dms,
dm_stats_counter_t counter
uint64_t region_id, uint64_t area_id);
int dm_stats_get_metric(const struct dm_stats *dms, int metric,
uint64_t region_id, uint64_t area_id,
double *value);
This simplifies the implementation of value aggregation for
groups of regions. The named function interface now calls the
enum interface internally so that all new functionality is
available regardless of the method used to retrieve values.
Cache the device-mapper name of a bound device in the dm_stats
handle.
This will be used by stats groups to report a device name or
user defined alias for groups.
The device-mapper name, device numbers and uuid stored in the
dm_stats handle are used only to bind the handle to a specific
device in order to issue ioctls.
Rename them to "bind_*" to reflect this usage in preparation
for caching the device-mapper name of the bound device in the
dm_stats handle.
This will be used to allow optional aliases to be set for
dmstats groups.
Check that @stats_list and @stats_print returned data in the
_stats_parse_list() and _stats_parse_region() functions before
attempting to operate on region and area values.
This avoids a coverity warning since fgets() could potentially
return no data from the memory buffer returned by the ioctl.
In both cases the ioctl would return an error, preventing these
functions from running, however it is cleaner to test for the
condition explicitly and fail in those cases.
Split up _build_histogram_arg() into separate functions to allocate
and fill the histogram arg string and remove nested local variable
declarations from the parent function.
Coverity flags a user-after-free in _stats_histograms_destroy():
>>> Calling "dm_pool_free" frees pointer "mem->chunk" which has
>>> already been freed.
This should not be possible since the histograms are destroyed in
reverse order of allocation:
203 for (n = _nr_areas_region(region) - 1; n; n--)
204 if (region->counters[n].histogram)
205 dm_pool_free(mem, region->counters[n].histogram);
It appears that Coverity is unaware that pool->chunk is updated
during the call to dm_pool_free() and valgrind flags no errors in
this function when called with multiple allocated histograms.
Since there is no actual need to free the histograms individually
in this way simplify the code and just free the first allocated
object (which will also free all later allocated histograms in a
single call).
The histogram changes adds a new error path to dm_stats_create().
Make sure that the dm_stats handle is properly destroyed if we fail
to create the histogram pool and check for failures setting the
program_id.
Since we are growing an object in the histogram pool the return
value of dm_pool_grow_object() must be checked and error paths need
to abandon the object before returning.
Older versions of gcc aren't able to track the assignments of
local variables as well as the latest versions leading to spurious
warnings like:
libdm-stats.c:2183: warning: "len" may be used uninitialized in this
function
libdm-stats.c:2177: warning: "minwidth" may be used uninitialized in
this function
Both of these variables are in fact assigned in all possible paths
through the function and later compilers do not produce these
warnings.
There's no reason to not initialize these variables though and
it makes the function slightly easier to follow.
Also fix one use of 'unsigned' for a nr_bins value.
Add support for creating, parsing, and reporting dm-stats latency
histograms on kernels that support precise_timestamps.
Histograms are specified as a series of time values that give the
boundaries of the bins into which I/O counts accumulate (with
implicit lower and upper bounds on the first and last bins).
A new type, struct dm_histogram, is introduced to represent
histogram values and bin boundaries.
The boundary values may be given as either a string of values (with
optional unit suffixes) or as a zero terminated array of uint64_t
values expressing boundary times in nanoseconds.
A new bounds argument is added to dm_stats_create_region() which
accepts a pointer to a struct dm_histogram initialised with bounds
values.
Histogram data associated with a region is parsed during a call to
dm_stats_populate() and used to build a table of histogram values
that are pointed to from the containing area's counter set. The
histogram for a specified area may then be obtained and interogated
for values and properties.
This relies on kernel support to provide the boundary values in
a @stats_list response: this will be present in 4.3 and 4.2-stable. A
check for a minimum driver version of 4.33.0 is implemented to ensure
that this is present (4.32.0 has the necessary precise_timestamps and
histogram features but is unable to report these via @stats_list).
Access methods are provided to retrieve histogram values and bounds
as well as simple string representations of the counts and bin
boundaries. Methods are also available to return the total count
for a histogram and the relative value (as a dm_percent_t) of a
specified bin.
Add support for the kernel precise_timestamps feature. This allows
regions to be created using counters with nanosecond precision.
A new dm_stats method, dm_stats_set_precise_timestamps() causes all
future regions created with this handle to attempt to enable precise
counters.
Several interfaced in libdm-stats return a uint64_t when it is
only used to signal success/failure: change all these uses to
return a simple int instead.
Add a method to retrieve the offset of an area within the
containing region (rather than the offset within the containing
device returned by dm_stats_get_area_start()).
Although users of the library can calculate this themselves it is
better to provide this through a method call to avoid users making
assumptions about the structure of regions and areas.
The dm_stats_get_area_start (and its '_current_' variant) methods
are expected to return the start sector of the area in the
containing device.
Make sure the call adds region->start to the returned value.
Since libdm-stats only uses fmemopen'd FILE objects the only way
that a close can fail is corruption of the memory containing the
FILE: check for this case and emit a backtrace if it occurs.
libdm/libdm-stats.c: 338 in _stats_parse_list()
libdm/libdm-stats.c: 341 in _stats_parse_list()
libdm/libdm-stats.c: 481 in _stats_parse_region()
libdm/libdm-stats.c: 487 in _stats_parse_region()
libdm/libdm-stats.c: 487 in _stats_parse_region()
- Calling "fclose" without checking return value
Remove an unneccessary conditional operator and simplify the logic
in _nr_areas:
libdm/libdm-stats.c: 501 in _nr_areas() - Control flow issues (DEADCODE)
Make sure the newly created handle is freed if we are unable to
also create the pool for it.
tools/dmsetup.c: 4255 in _stats_list() - Variable "dms" going out of scope leaks the storage it points to.
Make sure comm is closed in the error path of _program_id_from_proc().
libdm/libdm-stats.c: 98 in dm_stats_create() - Variable "comm" going out of scope leaks the storage it points to.
Add the libdm-stats module to libdm: this implements a simple interface
for creating, managing and interrogating I/O statistics regions and
areas on device-mapper devices.
The library interface is documented in libdevmapper.h and provides a
'dm_stats' handle that is used to perform statistics operations and
obtain data.
Public methods are provided to create and destroy handles and to list,
create, and destroy statistics regions as well as to obtain and parse
counter data and calculate rate-based metrics.
This commit also adds a 'dmsetup stats' (aka 'dmstats') command with
'clear', 'create', 'delete', 'list', 'print', and 'report' sub-commands.
See the library documentation and the dmstats.8 manual page for detailed
API and command descriptions.
