[igt-dev] [PATCH i-g-t 4/4] tools/intel_gpu_top: Add support of discrete GPU and card selection
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Thu May 28 14:28:51 UTC 2020
On 26/05/2020 13:24, Arkadiusz Hiler wrote:
> From: Ayaz A Siddiqui <ayaz.siddiqui at intel.com>
>
> In intel_gpu_top device path was hard coded for integrated GPU.
>
> With this patch we:
> * use igt_device_scan library for device scanning,
> * make discrete GPU the default one,
> * provided options for card selection.
>
> Signed-off-by: Ayaz A Siddiqui <ayaz.siddiqui at intel.com>
> Signed-off-by: Arkadiusz Hiler <arkadiusz.hiler at intel.com>
> ---
> man/intel_gpu_top.rst | 29 +++++++++-
> tools/Makefile.am | 2 +-
> tools/intel_gpu_top.c | 127 ++++++++++++++++++++++++++++++++++++------
> tools/meson.build | 2 +-
> 4 files changed, 140 insertions(+), 20 deletions(-)
>
> diff --git a/man/intel_gpu_top.rst b/man/intel_gpu_top.rst
> index d487baca..5552e969 100644
> --- a/man/intel_gpu_top.rst
> +++ b/man/intel_gpu_top.rst
> @@ -7,9 +7,9 @@ Display a top-like summary of Intel GPU usage
> ---------------------------------------------
> .. include:: defs.rst
> :Author: IGT Developers <igt-dev at lists.freedesktop.org>
> -:Date: 2019-02-08
> +:Date: 2020-03-18
> :Version: |PACKAGE_STRING|
> -:Copyright: 2009,2011,2012,2016,2018,2019 Intel Corporation
> +:Copyright: 2009,2011,2012,2016,2018,2019,2020 Intel Corporation
> :Manual section: |MANUAL_SECTION|
> :Manual group: |MANUAL_GROUP|
>
> @@ -43,6 +43,31 @@ OPTIONS
>
> -s <ms>
> Refresh period in milliseconds.
> +-L
> + List available GPUs on the platform.
> +-d
> + Select a specific GPU using supported filter.
> +
> +
> +DEVICE SELECTION
> +================
> +
> +User can select specific GPU for performance monitoring on platform where multiple GPUs are available.
> +A GPU can be selected by sysfs path, drm node or using various PCI sub filters.
> +
> +Filter types: ::
> +
> + ---
> + filter syntax
> + ---
> + sys sys:/sys/devices/pci0000:00/0000:00:02.0
> + find device by its sysfs path
> +
> + drm drm:/dev/dri/* path
> + find drm device by /dev/dri/* node
> +
> + pci pci:[vendor=%04x/name][,device=%04x][,card=%d]
> + vendor is hex number or vendor name
>
> LIMITATIONS
> ===========
> diff --git a/tools/Makefile.am b/tools/Makefile.am
> index 9352b41c..f97f9e08 100644
> --- a/tools/Makefile.am
> +++ b/tools/Makefile.am
> @@ -24,4 +24,4 @@ AM_CFLAGS = $(DEBUG_CFLAGS) $(DRM_CFLAGS) $(PCIACCESS_CFLAGS) $(CWARNFLAGS) \
> LDADD = $(top_builddir)/lib/libintel_tools.la
> AM_LDFLAGS = -Wl,--as-needed
>
> -intel_gpu_top_LDADD = $(top_builddir)/lib/libigt_perf.la
> +intel_gpu_top_LDADD = $(top_builddir)/lib/libigt_perf.la $(top_builddir)/lib/libigt_device_scan.la $(LIBUDEV_LIBS) $(GLIB_LIBS)
> diff --git a/tools/intel_gpu_top.c b/tools/intel_gpu_top.c
> index 8197482d..b3bbe505 100644
> --- a/tools/intel_gpu_top.c
> +++ b/tools/intel_gpu_top.c
> @@ -22,6 +22,7 @@
> */
>
> #include <stdio.h>
> +#include "igt_device_scan.h"
> #include <sys/types.h>
> #include <dirent.h>
> #include <stdint.h>
> @@ -94,7 +95,16 @@ struct engines {
> struct pmu_counter imc_reads;
> struct pmu_counter imc_writes;
>
> + bool discrete;
> + char *device;
> +
> + /* Do not edit below this line.
> + * This structure is reallocated every time a new engine is
> + * found and size is increased by sizeof (engine).
> + */
> +
> struct engine engine;
> +
> };
>
> static uint64_t
> @@ -168,14 +178,20 @@ static int engine_cmp(const void *__a, const void *__b)
> return a->instance - b->instance;
> }
>
> -static struct engines *discover_engines(void)
> +#define is_igpu_pci(x) (strcmp(x, "0000:00:02.0") == 0)
> +#define is_igpu(x) (strcmp(x, "i915") == 0)
> +
> +static struct engines *discover_engines(char *device)
> {
> - const char *sysfs_root = "/sys/devices/i915/events";
> + char sysfs_root[PATH_MAX];
> struct engines *engines;
> struct dirent *dent;
> int ret = 0;
> DIR *d;
>
> + snprintf(sysfs_root, sizeof(sysfs_root),
> + "/sys/devices/%s/events", device);
> +
> engines = malloc(sizeof(struct engines));
> if (!engines)
> return NULL;
> @@ -183,6 +199,8 @@ static struct engines *discover_engines(void)
> memset(engines, 0, sizeof(*engines));
>
> engines->num_engines = 0;
> + engines->device = device;
> + engines->discrete = !is_igpu(device);
>
> d = opendir(sysfs_root);
> if (!d)
> @@ -497,7 +515,7 @@ static int pmu_init(struct engines *engines)
> }
>
> engines->rapl_fd = -1;
> - if (rapl_type_id()) {
> + if (!engines->discrete && rapl_type_id()) {
> engines->rapl_scale = rapl_gpu_power_scale();
> engines->rapl_unit = rapl_gpu_power_unit();
> if (!engines->rapl_unit)
> @@ -695,8 +713,11 @@ usage(const char *appname)
> "\t[-l] List plain text data.\n"
> "\t[-o <file|->] Output to specified file or '-' for standard out.\n"
> "\t[-s <ms>] Refresh period in milliseconds (default %ums).\n"
> + "\t[-L] List all cards.\n"
> + "\t[-d <device>] Device filter, please check manual page for more details.\n"
> "\n",
> appname, DEFAULT_PERIOD_MS);
> + igt_device_print_filter_types();
> }
>
> static enum {
> @@ -1082,13 +1103,18 @@ print_header(struct engines *engines, double t,
> if (output_mode == INTERACTIVE) {
> printf("\033[H\033[J");
>
> - if (lines++ < con_h)
> - printf("intel-gpu-top - %s/%s MHz; %s%% RC6; %s %s; %s irqs/s\n",
> - freq_items[1].buf, freq_items[0].buf,
> - rc6_items[0].buf, power_items[0].buf,
> - engines->rapl_unit,
> - irq_items[0].buf);
> -
> + if (lines++ < con_h) {
> + if (!engines->discrete)
> + printf("intel-gpu-top - %s/%s MHz; %s%% RC6; %s %s; %s irqs/s\n",
> + freq_items[1].buf, freq_items[0].buf,
> + rc6_items[0].buf, power_items[0].buf,
> + engines->rapl_unit,
> + irq_items[0].buf);
> + else
> + printf("intel-gpu-top - %s/%s MHz; %s%% RC6; %s irqs/s\n",
> + freq_items[1].buf, freq_items[0].buf,
> + rc6_items[0].buf, irq_items[0].buf);
> + }
> if (lines++ < con_h)
> printf("\n");
> }
> @@ -1249,6 +1275,33 @@ static void sigint_handler(int sig)
> stop_top = true;
> }
>
> +/* tr_pmu_name()
> + *
> + * Transliterate pci_slot_id to sysfs device name entry for discrete GPU.
> + * Discrete GPU PCI ID ("xxxx:yy:zz.z") device = "i915_xxxx_yy_zz.z".
> + */
> +static char *tr_pmu_name(struct igt_device_card *card)
> +{
> + int ret;
> + const int bufsize = 18;
> + char *buf, *device = NULL;
> +
> + assert(card->pci_slot_name[0]);
> +
> + device = malloc(bufsize);
> + assert(device);
> +
> + ret = snprintf(device, bufsize, "i915_%s", card->pci_slot_name);
> + assert(ret == (bufsize-1));
> +
> + buf = device;
> + for (; *buf; buf++)
> + if (*buf == ':')
> + *buf = '_';
> +
> + return device;
> +}
> +
> int main(int argc, char **argv)
> {
> unsigned int period_us = DEFAULT_PERIOD_MS * 1000;
> @@ -1256,10 +1309,14 @@ int main(int argc, char **argv)
> char *output_path = NULL;
> struct engines *engines;
> unsigned int i;
> - int ret, ch;
> + int ret = 0, ch;
> + bool list_device = false;
> + enum igt_devices_print_type printtype = IGT_PRINT_SIMPLE;
> + char *pmu_device, *opt_device = NULL;
> + struct igt_device_card card;
>
> /* Parse options */
> - while ((ch = getopt(argc, argv, "o:s:Jlh")) != -1) {
> + while ((ch = getopt(argc, argv, "o:s:d:JLlh")) != -1) {
> switch (ch) {
> case 'o':
> output_path = optarg;
> @@ -1267,9 +1324,15 @@ int main(int argc, char **argv)
> case 's':
> period_us = atoi(optarg) * 1000;
> break;
> + case 'd':
> + opt_device = strdup(optarg);
> + break;
> case 'J':
> output_mode = JSON;
> break;
> + case 'L':
> + list_device = true;
> + break;
> case 'l':
> output_mode = STDOUT;
> break;
> @@ -1320,19 +1383,46 @@ int main(int argc, char **argv)
> break;
> };
>
> - engines = discover_engines();
> + igt_devices_scan(false);
> +
> + if (list_device) {
> + igt_devices_print(printtype);
I am not really happy with the output here, but that criticism goes to
the lsgpu and library as well. It seems that we are listing multiple
entries per card (not talking about render nodes!) so it is trial and
error to copy & paste the right one for giving to "-D".
My opens:
Should lsgpu list master and render nodes separately? I would say not by
default, or at least list them in a hierarchical manner akin to lsblk.
Should the output be condensed to one master entry per physical GPU by
default? (And be the one which works when passed in to "-D". Not all do,
because some don't have the PCI data, while they do point to the same
DRM card, AFAIR.
> + goto exit;
> + }
> +
> + if (opt_device != NULL) {
> + /* Free opt device now as its not needed further. */
> + ret = igt_device_card_match(opt_device, &card);
> + free(opt_device);
> + if (!ret) {
> + fprintf(stderr, "Requested device %s not found!\n", opt_device);
opt_device is use after free here.
> + ret = 1;
Instead these ones (here and below) we could use EXIT_FAILURE for
readability.
> + goto exit;
> + }
> + } else {
> + igt_device_find_first_discrete_card(&card);
> + }
> +
> + if (card.pci_slot_name[0] && !is_igpu_pci(card.pci_slot_name))
> + pmu_device = tr_pmu_name(&card);
> + else
> + pmu_device = strdup("i915");
> +
> + engines = discover_engines(pmu_device);
> if (!engines) {
> fprintf(stderr,
> "Failed to detect engines! (%s)\n(Kernel 4.16 or newer is required for i915 PMU support.)\n",
> strerror(errno));
> - return 1;
> + ret = 1;
> + goto err;
> }
>
> ret = pmu_init(engines);
> if (ret) {
> fprintf(stderr,
> "Failed to initialize PMU! (%s)\n", strerror(errno));
> - return 1;
> + ret = 1;
> + goto err;
> }
>
> pmu_sample(engines);
> @@ -1384,5 +1474,10 @@ int main(int argc, char **argv)
> usleep(period_us);
> }
>
> - return 0;
> +err:
> + free(engines);
> + free(pmu_device);
> +exit:
> + igt_devices_free();
> + return ret;
> }
> diff --git a/tools/meson.build b/tools/meson.build
> index 59b56d5d..34f95e79 100644
> --- a/tools/meson.build
> +++ b/tools/meson.build
> @@ -93,7 +93,7 @@ install_subdir('registers', install_dir : datadir,
> executable('intel_gpu_top', 'intel_gpu_top.c',
> install : true,
> install_rpath : bindir_rpathdir,
> - dependencies : lib_igt_perf)
> + dependencies : [lib_igt_perf,lib_igt_device_scan])
>
> executable('amd_hdmi_compliance', 'amd_hdmi_compliance.c',
> dependencies : [tool_deps],
>
Apart from two minor nits and a generic open on lspgu behaviour I tested
this and it worked.
Regards,
Tvrtko
More information about the igt-dev
mailing list