Skip to content

Commit

Permalink
host_exerciser:support no FPGA mgmt PF instances (#3108) (#3109)
Browse files Browse the repository at this point in the history 
* host_exerciser:support no FPGA mgmt PF instances

Issue: The host exerciser fails to execute on Devkits if it does not support FPGA management PF.
The host exerciser enumerates FME and quits the program if the bitstream does not support it.

The I series-DK and F series-DK have multiple host exerciser AFUs on separate PCIe bus, device functions.

fix: If the bitstream does not support FME, perform the host exerciser tests and print message.

host_exerciser --pci-address 0000:03:00.2 lpbk
host_exerciser --pci-address 0000:04:00.2 lpbk

opae.io ls
[0000:03:00.0] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:04:00.7] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: None)
[0000:04:00.5] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)
[0000:03:00.7] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)
[0000:04:00.3] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:03:00.5] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)
[0000:04:00.1] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:03:00.3] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:03:00.1] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:04:00.6] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: None)
[0000:04:00.4] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:03:00.6] (0x8086:0xbccf 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)
[0000:04:00.2] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)
[0000:03:00.4] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:04:00.0] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: dfl-pci)
[0000:03:00.2] (0x8086:0xbcce 0x8086:0x1771) Intel Acceleration Development Platform N6001 (Driver: vfio-pci)

--------

Signed-off-by: anandaravuri <ananda.ravuri@intel.com>
Co-authored-by: Peter Colberg <peter.colberg@intel.com>
  • Loading branch information
@anandaravuri @pcolberg
anandaravuri and pcolberg committed Feb 26, 2024
1 parent 5447d98 commit e9e9ac8
Show file tree
Hide file tree
Showing 3 changed files with 92 additions and 72 deletions.
72 changes: 40 additions & 32 deletions libraries/afu-test/afu_test.h
View file
Expand Up @@ -172,6 +172,7 @@ class afu {
, shared_(false)
, timeout_msec_(60000)
, handle_(nullptr)
, handle_device_(nullptr)
, current_command_(nullptr)
{
if (!afu_id_.empty())
Expand All @@ -196,55 +197,65 @@ class afu {
return fpga::properties::get(handle_);
}

int open_handle(const char *afu_id) {
bool enum_fpga_device()
{
auto filter = fpga::properties::get(); // Get an empty properties object

// The following code attempts to get a token+handle for the DEVICE.
// The following code attempts to get a token+handle for the DEVICE.
// This is to allow access to OPAE-API functions that are only supported
// through the xfpga plugin (i.e accessing sysfs entries)
// In contrast, the ACCELERATOR token may be underlied by the vfio plugin.
// Set PCIe segment, bus, and device properties to enumerate FPGA DEVICE (FME)
if (!pci_addr_.empty()) {
auto p = pcie_address::parse(pci_addr_.c_str());
filter->segment = p.fields.domain;
filter->bus = p.fields.bus;
filter->device = p.fields.device;
auto p = pcie_address::parse(pci_addr_.c_str());
filter->segment = p.fields.domain;
filter->bus = p.fields.bus;
filter->device = p.fields.device;
}

filter->type = FPGA_DEVICE;
auto tokens = fpga::token::enumerate({filter});
auto tokens = fpga::token::enumerate({ filter });

// Error out if the # of tokens != 1
if (tokens.size() < 1) {
if (pci_addr_.empty()) {
logger_->error("no DEVICE found");
} else {
logger_->error("no accelerator found at PCIe address {1}",
pci_addr_);
}
return exit_codes::not_found;
}
logger_->info("no FPGA DEVICE found");
return false;
}

if (tokens.size() > 1) {
std::cerr << "more than one DEVICE found matching filter\n";
logger_->info("more than one FPGA DEVICE found ");
return false;
}

int flags = shared_ ? FPGA_OPEN_SHARED : 0;

// Open a handle to the resource
try {
handle_device_ = fpga::handle::open(tokens[0], flags);
} catch (fpga::no_access &err) {
std::cerr << err.what() << "\n";
return exit_codes::no_access;
handle_device_ = fpga::handle::open(tokens[0], flags);
}
catch (fpga::no_access& err) {
std::cerr << err.what() << "\n";
return false;
}
return true;
}

// The following code attempts to get a token + handle for the AFU
// (ACCELERATOR device) matching the given command's afu_id.
// Set PCIe segment, bus, device, and functionproperties to enumerate FPGA ACCELERATOR
int open_handle(const char *afu_id) {

enum_fpga_device();

auto filter = fpga::properties::get(); // Get an empty properties object

// The following code attempts to get a token+handle for the DEVICE.
// This is to allow access to OPAE-API functions that are only supported
// through the xfpga plugin (i.e accessing sysfs entries)
// In contrast, the ACCELERATOR token may be underlied by the vfio plugin.
// Set PCIe segment, bus, and device properties to enumerate FPGA DEVICE (FME)
if (!pci_addr_.empty()) {
auto p = pcie_address::parse(pci_addr_.c_str());
filter->function = p.fields.function;
auto p = pcie_address::parse(pci_addr_.c_str());
filter->segment = p.fields.domain;
filter->bus = p.fields.bus;
filter->device = p.fields.device;
filter->function = p.fields.function;
}

auto app_afu_id = afu_id ? afu_id : afu_id_.c_str();
Expand All @@ -255,7 +266,7 @@ class afu {
return error;
}

tokens = fpga::token::enumerate({filter});
auto tokens = fpga::token::enumerate({filter});
if (tokens.size() < 1) {
if (pci_addr_.empty()) {
logger_->error("no accelerator found with id: {0}", app_afu_id);
Expand All @@ -266,10 +277,7 @@ class afu {
return exit_codes::not_found;
}

if (tokens.size() > 1) {
std::cerr << "more than one accelerator found matching filter\n";
}

int flags = shared_ ? FPGA_OPEN_SHARED : 0;
try {
handle_ = fpga::handle::open(tokens[0], flags);
} catch (fpga::no_access &err) {
Expand Down
4 changes: 3 additions & 1 deletion samples/host_exerciser/host_exerciser.h
View file
Expand Up @@ -575,7 +575,9 @@ class host_exerciser : public test_afu {

token::ptr_t get_token_device()
{
return handle_device_->get_token();
if (handle_device_)
return handle_device_->get_token();
return nullptr;
}

bool option_passed(std::string option_str)
Expand Down
88 changes: 49 additions & 39 deletions samples/host_exerciser/host_exerciser_cmd.h
View file
Expand Up @@ -737,46 +737,8 @@ class host_exerciser_cmd : public test_command
host_exe_ = dynamic_cast<host_exerciser*>(afu);

token_ = d_afu->get_token();
token_device_ = d_afu->get_token_device();

// Check if memory calibration has failed and error out before proceeding
// with the test. The dfl-emif driver creates sysfs entries to report the
// calibration status for each memory channel. sysobjects are the OPAE-API's
// abstraction for sysfs entries. However, at this time, these are only
// accessible through tokens that use the xfpga plugin and not the vfio
// plugin. Hence our use of the DEVICE token (token_device_). One
// non-ideality of the following implementation is the use of
// MAX_NUM_MEM_CHANNELS. We are essentially doing a brute-force query of
// sysfs entries since we don't know how many mem channels exist on the
// given platform. What about glob wildcards? Why not simply glob for
// "*dfl*/**/inf*_cal_fail" and use the OPAE-API's support for arrays of
// sysobjects? The reason is that, at the time of this writing, the
// xfpga-plugin's sysobject implementation does not support arrays
// specifically when the glob contains a recursive wildcard "/**/". It's a
// strange and perhaps unnecessary limitation. Therefore, future work is to
// fix that and clean up the code below.
for (size_t i = 0; i < MAX_NUM_MEM_CHANNELS; i++) {
std::stringstream mem_cal_glob;
// Construct the glob string to search for the cal_fail sysfs entry
// for the i'th mem channel
mem_cal_glob << "*dfl*/**/inf" << i << "_cal_fail";
// Ask for a sysobject with this glob string
fpga::sysobject::ptr_t testobj = fpga::sysobject::get(
token_device_, mem_cal_glob.str().c_str(), FPGA_OBJECT_GLOB);

// if test obj !=null, the sysfs entry was found.
// Read the calibration status from the sysfs entry.
// A non-zero value (typically '1') means
// calibration has failed --> we error out.
if (testobj && testobj->read64(0)) {
std::cout
<< "This sysfs entry reports that memory calibration has failed:"
<< mem_cal_glob.str().c_str() << std::endl;
return -1;
}
}


fpga_emif_status(afu);
// Read HW details
uint64_t he_info = host_exe_->read64(HE_INFO0);
he_lpbk_api_ver_ = (he_info >> 16);
Expand Down Expand Up @@ -880,6 +842,54 @@ class host_exerciser_cmd : public test_command

return status;
}

void fpga_emif_status(test_afu* afu)
{
auto d_afu = dynamic_cast<host_exerciser*>(afu);
token_device_ = d_afu->get_token_device();

if (!token_device_)
return;

// Check if memory calibration has failed and error out before proceeding
// with the test. The dfl-emif driver creates sysfs entries to report the
// calibration status for each memory channel. sysobjects are the OPAE-API's
// abstraction for sysfs entries. However, at this time, these are only
// accessible through tokens that use the xfpga plugin and not the vfio
// plugin. Hence our use of the DEVICE token (token_device_). One
// non-ideality of the following implementation is the use of
// MAX_NUM_MEM_CHANNELS. We are essentially doing a brute-force query of
// sysfs entries since we don't know how many mem channels exist on the
// given platform. What about glob wildcards? Why not simply glob for
// "*dfl*/**/inf*_cal_fail" and use the OPAE-API's support for arrays of
// sysobjects? The reason is that, at the time of this writing, the
// xfpga-plugin's sysobject implementation does not support arrays
// specifically when the glob contains a recursive wildcard "/**/". It's a
// strange and perhaps unnecessary limitation. Therefore, future work is to
// fix that and clean up the code below.

for (size_t i = 0; i < MAX_NUM_MEM_CHANNELS; i++) {
std::stringstream mem_cal_glob;
// Construct the glob string to search for the cal_fail sysfs entry
// for the i'th mem channel
mem_cal_glob << "*dfl*/**/inf" << i << "_cal_fail";
// Ask for a sysobject with this glob string
fpga::sysobject::ptr_t testobj = fpga::sysobject::get(
token_device_, mem_cal_glob.str().c_str(), FPGA_OBJECT_GLOB);

// if test obj !=null, the sysfs entry was found.
// Read the calibration status from the sysfs entry.
// A non-zero value (typically '1') means
// calibration has failed --> we error out.
if (testobj && testobj->read64(0)) {
std::cout
<< "This sysfs entry reports that memory calibration has failed:"
<< mem_cal_glob.str().c_str() << std::endl;
return;
}
}

}

protected:
he_cfg he_lpbk_cfg_;
Expand Down

0 comments on commit e9e9ac8

Please sign in to comment.