Hi dude,
Another suggestions, when you release this, can you let us know which
files, if any, need to be changed in the asm/perfmon.h or
arch/xxx/perfmon directory? Also, is there a way you can provide a
'diff' against a previously patched kernel? That would be TREMENDOUSLY
helpful.
Phil
P.S. Sorry about my griping before...I'm overloaded. If it helps, I'm
poking around on a Cell machine looking to attempt a perfmon patch at
the moment.
On Wed, 2006-11-01 at 12:53 -0800, Stephane Eranian wrote:
> Hello everyone,
>
> I have made the final round of changes to the interface to address
> some issues reported during the review on LKML.
>
> Unfortunately, those changes are NOT backward compatible with existing 2.2
> interface.
> As a consequence, I have now changed the version of the perfmon development
> source to
> 2.3 (from 2.2). Backward compatibility with existing mainline (and distro)
> perfmon
> for Itanium is MAINTAINED AND UNCHANGED.
>
> The changes are as follows:
>
> 1/ pfm_create_evtsets()/pfm_delete_evtsets()
>
> The pfarg_setdesc_t structure has changed:
> typedef struct {
> uint16_t set_id; /* which set */
> uint16_t set_id_next; /* next set to go to (must use
> PFM_SETFL_EXPL_NEXT) */
> uint32_t set_flags; /* input: flags for set, output:
> err flag */
> uint64_t set_timeout; /* requested/effective switch
> timeout in nsecs */
> ^^^^^^^^
> uint64_t set_mmap_offset; /* cookie to pass as mmap offset to
> access 64-bit virtual PMD */
> uint64_t reserved[5]; /* for future use */
> } pfarg_setdesc_t;
>
> The set_timeout is now 64-bit wide and is expressed in nanosecond. This
> is more
> or less how time (or timeout) is expressed with other system calls (e.g.,
> nanosleep, gettimeofday).
>
> 2/ pfm_getinfo_evtsets()
>
> The pfarg_setinfo_t structure has changed to match the changes in
> pfarg_setdesc_t:
> typedef struct {
> uint16_t set_id; /* which set */
> uint16_t set_id_next; /* output: next set to go to
> (must use PFM_SETFL_EXPL_NEXT) */
> uint32_t set_flags; /* output:flags or error */
> uint64_t set_ovfl_pmds[PFM_PMD_BV]; /* output: last ovfl PMDs
> which triggered a switch from set */
> uint64_t set_runs; /* output: number of times the
> set was active */
> uint64_t set_timeout; /* output:effective/leftover
> switch timeout in nsecs */
> ^^^^^^^^
> uint64_t set_act_duration; /* DO NOT USE: number of cycles
> set was active (syswide only) */
> uint64_t set_mmap_offset; /* cookie to pass as mmap offset
> to access 64-bit virtual PMD */
> uint64_t set_avail_pmcs[PFM_PMC_BV];
> uint64_t set_avail_pmds[PFM_PMD_BV];
> uint64_t reserved[4]; /* for future use */
> }
>
> 3/ UUID removal
>
> Sampling formats are no longer identified by 128-bit UUID. People thought
> it was overkill and
> difficult to manipulate. They suggested using simple strings which are
> much easier to manipulate.
> I made the change now.
>
> Here are the mappings:
> perfmon v2.0 default format -> "default-old"
> perfmon v2.2 v2.3 default format -> "default"
>
> When operating in v2.0 compatibility mode (i.e., using perfmonctl()),
> there is no need to remove the
> UUID parameter. The kernel will convert it to "default-old". Only the old
> default format will be accepted
> when coming via perfmonctl().
>
> The information reported in /sys/kernel/perfmon/formats has changed
> accordingly. The clear text UUID is gone.
> Each format gets a directory (as before), and inside each directory you
> find 2 entries:
> - smpl_version: print a version number of rot eh format, expressed as
> X.Y
> - smpl_argsize: the size of the argument to the format for
> pfm_create_context()
>
> 4/ pfm_create_context()
>
> Given the changes with UUID, I have decided it was time to refresh the
> call completely to align it
> with other common Linux(or UNIX) practices. In particular, the following
> changes were made:
>
> - the file descriptor is now returned by the call rather than in the
> parameters. This follows the model
> of open(), socket().
>
> - the sampling format name is added as a parameter
>
> - the actual sampling buffer size is not returned anymore. For formats,
> that need a size, it is typically passed in
> the sampling format argument. You simply use the same size value for
> mmap, munmap.
> Both mmap() and munmap() align to the next page boundary
> automatically.
>
> - the pfarg_ctx_t structure has been vastly simplified to only contain
> ctx_flags. it is not modified by the call as it
> used to be in v2.2.
>
> Here is the new definition:
>
> int pfm_create_context(pfarg_ctx_t *ctx, char *fmt_name, char *fmt_arg,
> size_t fmt_argsize);
>
> Where pfarg_ctx_t is defined as:
>
> typedef struct {
> uint32_t ctx_flags; /* noblock/block/syswide */
> uint32_t ctx_reserved1; /* for future use */
> uint64_t ctx_reserved3[7]; /* for future use */
> } pfarg_ctx_t;
>
> fmt_name : the name of the format you want to use. Pass NULL if no
> format is used
> fmt_arg : pointer to argument for format. Not all formats take
> arguments, but the default does (no changes to it)
> fmt_arg_size: the size of the structure passed in fmt_arg
>
> Return:
> >=0 : the file desriptor for the context
> -1: error, errno contains the error code
>
> Internally as a consequence of the changes, the code has been simplified
> some more, especially for
> pfm_create_context(). So I think this is a change for the better.
>
> The changes will be part of the next kernel patch. Similarly, the next
> releases of pfmon/libpfm will
> adjust to the kernel changes.
>
> Any comments?
>
> --
> -Stephane
>
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
