| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
When using aggregation with http2, :authority headers were getting
forwarded to the client, which didn't know how to deal with them on
http1.
Filter all http2 headers.
Tested: Unit tests pass.
Change-Id: I6a834656b604004eeba1a2aa2f245ef211f28495
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Ignores any json property named `DateTime` when calculating
the etag value of an HTTP response as per the updated
Redfish Spec (section 6.5: ETags)
Tested:
- Redfish Service Validator passes
- Tested on romulus:
1. GET resource with a "DateTime" field
```
curl -k -H "X-Auth-Token: $XAUTH_TOKEN" "https://$BMC/redfish/v1/TaskService" \
--etag-save ./etag.txt -v
...
< etag: "6A4CE897"
...
{
"@odata.id": "/redfish/v1/TaskService",
"@odata.type": "#TaskService.v1_1_4.TaskService",
"CompletedTaskOverWritePolicy": "Oldest",
"DateTime": "2025-07-23T17:08:20+00:00",
"Id": "TaskService",
"LifeCycleEventOnTaskStateChange": true,
"Name": "Task Service",
"ServiceEnabled": true,
"Status": {
"State": "Enabled"
},
"Tasks": {
"@odata.id": "/redfish/v1/TaskService/Tasks"
}
```
2. GET same resource again later, etag is same as before
```
curl -k -H "X-Auth-Token: $XAUTH_TOKEN" "https://$BMC/redfish/v1/TaskService" \
--etag-save ./etag.txt -v
...
< etag: "6A4CE897"
...
{
"@odata.id": "/redfish/v1/TaskService",
"@odata.type": "#TaskService.v1_1_4.TaskService",
"CompletedTaskOverWritePolicy": "Oldest",
"DateTime": "2025-07-23T17:10:48+00:00",
"Id": "TaskService",
"LifeCycleEventOnTaskStateChange": true,
"Name": "Task Service",
"ServiceEnabled": true,
"Status": {
"State": "Enabled"
},
"Tasks": {
"@odata.id": "/redfish/v1/TaskService/Tasks"
}
```
"DateTime" is the only value to have changed, but since
it is ignored the etag did not change
3. GET with if-none-match returns 304
```
curl -k -H "X-Auth-Token: $XAUTH_TOKEN" "https://$BMC/redfish/v1/TaskService" \
--etag-save ./etag.txt --etag-compare ./etag.txt -v
...
> if-none-match: "6A4CE897"
...
< HTTP/2 304
< allow: GET
< odata-version: 4.0
< strict-transport-security: max-age=31536000; includeSubdomains
< pragma: no-cache
< cache-control: no-store, max-age=0
< x-content-type-options: nosniff
< etag: "6A4CE897"
< date: Wed, 23 Jul 2025 17:14:39 GMT
< content-length: 0
<
...
```
Change-Id: I51f7668e75719c69c55535e4a1e48c8bae7c9488
Signed-off-by: Corey Ethington <cethington@coreweave.com>
|
|
Our includes haven't been enforced by tidy in a while. Run the script,
check in the result, minus the false positives.
Change-Id: I6a6da26f5ba5082d9b4aa17cdc9f55ebd8cd41a6
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Within this namespace, we don't need to call crow, we are already in the
crow namespace.
Tested: Code compiles.
Change-Id: Ida57624ef1157f98f2719b5c3af536aebaca601e
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Resolve tracing of CRITICAL message on 1xx, 204(no-content) and 304(not
modified) when there is no body in the response.
The code is changed to check that 1xx, 204 and 304 response with
payloadsize of 0 will not trace the CRITICAL message.
Removed setting of no_content on PATCH that was being done before other
functions are called.
Tested: Used the following commands to check that the CRITICAL message
is no longer being traced for no-content and not modified response with
no body.
204 no-content
```
curl -k -H "Content-Type: application/json" \
-d '{"PowerRestorePolicy":"LastState"}' \
-X PATCH https://${bmc_ip}/redfish/v1/Systems/system
```
```
curl -k -H "Content-Type: application/json" \
-H "X-Auth-Token: $bmc_token" -X PATCH \
-d '{"LocationIndicatorActive":true}' \
https://${bmc_ip}/redfish/v1/Managers/bmc
```
304 not modified
```
curl -k -i https://${bmc_ip}/redfish/ --etag-save etag.out \
-H 'If-Modified-Since: Tue, 21 Nov 2050 08:00:00 GMT'
HTTP/1.1 200 OK
Allow: GET
.
.
.
ETag: "B3A9EAA1"
Content-Type: application/json
Date: Thu, 14 Aug 2025 22:07:27 GMT
Content-Length: 26
ETAG=`cat etag.out`; echo $ETAG
curl -k -i https://${bmc_ip}/redfish/ -H "If-None-Match: ${ETAG}"
HTTP/1.1 304 Not Modified
Allow: GET
.
.
.
ETag: "B3A9EAA1"
Date: Thu, 14 Aug 2025 22:16:27 GMT
Content-Length: 0
```
Change-Id: I98cc096c1f7e506687d4a6bf5a2e51b2231c0d68
Signed-off-by: Abiola Asojo <abiola.asojo@ibm.com>
|
|
Passing the TLS-provided credentials from the HTTP connection to the
http2 connection got missed, and appears to break mutual TLS for http2
connections. Pass the credentials.
Tested: Mutual TLS is now functional on http2 connections as shown in
the next patch.
Change-Id: Ia2bbcd5383dae859baa96908b76f221b9c74632c
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Trace URI on the existing CRITICAL message when a body exists but HTTP
return code was Informational responses (100 – 199), no-content (204),
or not_modified (304).
Tested: With these changes, the URI is traced on the CRITICAL message:
```
curl --http2 -k -H "Content-Type: application/json" \
-d '{"PowerRestorePolicy":"LastState"}' \
-X PATCH https://${bmc}/redfish/v1/Systems/system
root@p10bmc:~# journalctl | grep Response
Jun 20 15:30:46 p10bmc bmcwebd[296]: [http_response.hpp:213]
0x1353670 Response content provided but code was no-content
or not_modified, which aren't allowed to
have a body for url : "/redfish/v1/Systems/system"
```
Change-Id: I1ef618600642d355fc9f935d055b011e044caf5c
Signed-off-by: Abiola Asojo <abiola.asojo@ibm.com>
|
|
Given the size of Redfish schemas these days, it would be nice to be
able to store them on disk in a zstd format. Unfortunately, not all
clients support zstd at this time.
This commit implements reading of zstd files from disk, as well as
decompressing zstd in the case where the client does not support zstd as
a return type.
Tested:
Implanted an artificial zstd file into the system, and observed correct
decompression both with an allow-encoding header of empty string and
zstd.
Change-Id: I8b631bb943de99002fdd6745340aec010ee591ff
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Http2 support in bmcweb has been relatively stable for a while. The
http2 implementation passes all known Redfish tests (some of which
require ported to httpx to support http2), the UI loads, and so far as
the project is concerned, is a complete improvement over the existing
http1 stack.
This commit removes the experimental classification from http2, and
declares it ready for production use, while enabling it by default.
note, that enabling this by default only makes the server advertise that
http2 is available. Http2 must still be supported by the client to
enable ALPN negotiation, so existing http1 clients that only support
http1 will continue to function as they did before.
Tested: Enabled http option and saw http2 advertised, http2 now takes
effect.
Change-Id: I92843a3afc532f0b2a64904bb872e5d84a1a54fe
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
cppcheck found a case which dereferences an invalid iterator like
```
http/logging.hpp:74:12: warning: Either the condition 'it!=mapping.end()' is redundant or there is possible dereference of an invalid iterator: it. [derefInvalidIteratorRedundantCheck]
return it->second;
http/logging.hpp:69:12: note: Assuming that condition 'it!=mapping.end()' is not redundant
if (it != mapping.end())
^
http/logging.hpp:74:12: note: Dereference of an invalid iterator
return it->second;
^
```
Tested:
- Tries a various bmcweb loglevel.
Change-Id: Ieca8c5c5ee83f0b45a82c3d7e4f19b09bf1422e6
Signed-off-by: Myung Bae <myungbae@us.ibm.com>
|
|
All integers in a callback should be taken by value, not reference.
Fix a typo on "Async"
Implement a TODO on libssl, where a macro was removed, so we can now
directly call the function
Tested: Unit tests pass.
Change-Id: Iaccd100a6b5f0dc871ec2d2cb1f01bed3ceefde2
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
When websocket is closed due to a timeout under some situations like GUI
console not as responsive, itis currently logged as ERROR.
This commit changes it as WARNING.
```
May 02 10:09:31 ever28bmc pldmd[836]: BIOS attribute 'hb_cap_freq_mhz_min' updated to value '2000' by BMC 'false'
May 02 10:09:31 ever28bmc pldmd[836]: BIOS attribute 'hb_cap_freq_mhz_max' updated to value '3900' by BMC 'false'
May 02 10:09:31 ever28bmc bmcweb[1100]: [ERROR websocket.hpp:268] doRead error The socket was closed due to a timeout [boost.beast:1 at /usr/include/boost/beast/websocket/impl/stream_impl.hpp:346:13 in function 'bool boost::beast::websocket::stream< <template-parameter-1-1>, <anonymous> >::impl_type::check_stop_now(boost::beast::error_code&)']
May 02 10:09:34 ever28bmc pldmd[836]: BIOS attribute 'hb_max_number_huge_pages' updated to value '64' by BMC 'false'
```
Tested:
- Load GUI pages and kill web-browser multiple times while reading BMC
journal records.
Change-Id: I7e15845be7d3762ef144744ca1aedadf96e43a2f
Signed-off-by: Myung Bae <myungbae@us.ibm.com>
|
|
There are cases bmcweb might close the connection due to a violation of
the protocol. Currently these are done gracefully, under the assumption
that a client might attempt to recover. But this opens us up to
potentially leaving sockets open for far longer than we intend if the
client is completely gone, due to a disconnect or explicitly closing the
socket hard.
In cases where we get a protocol error, shutdown the socket hard, rather
than attempt to do things "correctly".
Tested:
I tested this MR using a script that simulated 5,000 parallel
connections simultaneously to BMC and closed them immediately without
properly sending a close_notify alert
Observations:
The BMC became unresponsive for 30-40 seconds before recovering.
After recovery, it took approximately 90 seconds to close all
connections in QEMU.
On real hardware, connection closure times may be slightly higher
(though still within expected parameters).
Conclusion:
This behavior aligns with expectations.
After 90 seconds observed that
1) No sockets in CLOSE_WAIT state
2) Able to make new connection.
```
curl -k -H "X-Auth-Token:$bmc_token" https://${IP}/redfish/v1/AccountService/Accounts
{
"@odata.id": "/redfish/v1/AccountService/Accounts",
"@odata.type": "#ManagerAccountCollection.ManagerAccountCollection",
"Description": "BMC User Accounts",
"Members": [
{
"@odata.id": "/redfish/v1/AccountService/Accounts/root"
}
],
"Members@odata.count": 1,
"Name": "Accounts Collection"
}
```
Change-Id: I6ab4347efd8fda9ae86bfbb8575666ad3eabe88c
Signed-off-by: Ed Tanous <etanous@nvidia.com>
Signed-off-by: Chandramohan Harkude <chandramohan.harkude@gmail.com>
|
|
Problem : When 201 connections made in parallel to BMC and closed them
immediately without properly sending a close_notify alert it was
observed that Bmcweb server was taking several minutes to close the
sockets. All the 200 TCP sockets were in CLOSE_WAIT state.
Journal log shows below line
[CRITICAL http_connection.hpp:213] 0x29d1ef0Max connection count
exceeded.
```
Not able to make new connection
$ curl -k -H "X-Auth-Token:$bmc_token" -X GET https://${BMC_IP}/redfish
/v1/AccountService/Accounts
curl: (35) OpenSSL SSL_connect: SSL_ERROR_SYSCALL in connection to
127.0.0.1:2443
Fix : The bmcweb server failed to identify the end of stream at TCP (SSL
/TLS) layer, therefore added check to identify the end of stream which
closes the connection and socket
Test :
I tested this MR using a script that simulated 5,000 parallel
connections simultaneously to BMC and closed them immediately without
properly sending a close_notify alert
Observations:
The BMC became unresponsive for 30-40 seconds before recovering.
After recovery, it took approximately 90 seconds to close all
connections in QEMU.
On real hardware, connection closure times may be slightly higher
(though still within expected parameters).
Conclusion:
This behavior aligns with expectations.
After 90 seconds observed that
1) No sockets in CLOSE_WAIT state
2) Able to make new connection.
curl -k -H "X-Auth-Token:$bmc_token" https://${IP}/redfish/v1/AccountService/Accounts
{
"@odata.id": "/redfish/v1/AccountService/Accounts",
"@odata.type": "#ManagerAccountCollection.ManagerAccountCollection",
"Description": "BMC User Accounts",
"Members": [
{
"@odata.id": "/redfish/v1/AccountService/Accounts/root"
}
],
"Members@odata.count": 1,
"Name": "Accounts Collection"
}
```
Change-Id: I1c277db0b774d33c656b4a2b1bd14f3575535bec
Signed-off-by: Chandramohan Harkude <chandramohan.harkude@gmail.com>
|
|
as we have successfully merged patches that enable UserPrincipalName
parse mode, we can start removing Meta only parse mode. This commit
is intended to remove MTLSCommonNameParseMode::Meta from the upstream
code
Tested:
- build bmcweb
- deploy to a device that already use UPN
- check if it works fine by sending curl request /AccountService
Change-Id: Idcf4340a2a9940f035aea41cd30ef4df7bd95530
Signed-off-by: Malik Akbar Hashemi Rafsanjani <malikrafsan@meta.com>
|
|
I don't feel like breaking these out at the moment or writing a commit
message. This fixes the build for clang-tidy. If anyone wants to break
these out with appropriate commit messages, feel free.
Change-Id: Id0b65d238dfb9b8036c0ffddf2f32d221e5988c2
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
This commit is fixing coverity issues reported for copy in stead of
move.
Tested: redfish service validator passes
Change-Id: I97e755830f28390e7c4bfaba6f3f947898a21423
Signed-off-by: Ed Tanous <ed@tanous.net>
|
|
The error log for connection failures previously displayed an invalid
host name and port number because it retrieved these values from the
endpoint argument in the afterConnect callback, which is not valid when
the connection fails. This patch resolves the issue by obtaining the
host name directly from the connection object.
Tested By:
Attempted to subscribe to an unreachable host and sent test events to
verify that the error log was generated as expected.
Change-Id: I33772b224eb7ab164c2de342f35ae42c34126b3d
Signed-off-by: Abhilash Raju <abhilash.kollam@gmail.com>
|
|
It is currently too easy to accidentally make copies of the Request
object. Ideally code would parse out the Request in the first handler,
then no longer require an async copy. There is one case in the redfish
query things where we actually need a copy of the request object, so we
need these constructors, but we should make them explicit.
This commit moves the Request constructor to be private, and adds a new
method called copy() for explicitly making a copy. Ironcially, this
finds one place where we were actually making a copy of the request
object unintentionally, so fix that to only capture the value
required,the user session.
Tested:
- Compiles
- Run GET/PATCH related curl or If-Match like PATCH Account
- Redfish Service Validator runs and passes
Change-Id: I19255981f42757ed736112c003201e3f758735ac
Signed-off-by: Ed Tanous <ed@tanous.net>
Signed-off-by: Myung Bae <myungbae@us.ibm.com>
|
|
Goal of the MR is to provide infrastructure support in bmcweb to manage
the OEM fragment handling separately. OEM schema are vendor defined and
per DMTF resource we could have multiple vendor defined OEM schema to be
enabled.
The feature allows registration of route handler per schema per OEM
namespace.
Example
```
REDFISH_SUB_ROUTE<"/redfish/v1/Managers/<str>/#/Oem/OpenBmc">(service,
HttpVerb::Get)(oemOpenBmcCallback);
REDFISH_SUB_ROUTE<"/redfish/v1/Managers/<str>/#/Oem/Nvidia">(service,
HttpVerb::Get)(oemNidiaCallback);
```
We can have separate vendor defined route handlers per resource. Each of
these route handlers can populate their own vendor specific OEM data.
The OEM code can be better organized and enabled/disabled as per the
platform needs. The current MR has the code changes related to handling
GET requests alone. The feature only supports requests
where the response payload is JSON.
Tests
- All UT cases passes
- New UT added for RF OEM router passes
- Service Validator passes on qemu
- GET Response on Manager/bmc resource contains the OEM fragment
```
curl -c cjar -b cjar -k -X GET https://127.0.0.1:2443/redfish/v1/Managers/bmc
{
"@odata.id": "/redfish/v1/Managers/bmc",
"@odata.type": "#Manager.v1_14_0.Manager",
"Oem": {
"OpenBmc": {
"@odata.id": "/redfish/v1/Managers/bmc#/Oem/OpenBmc",
"@odata.type": "#OpenBMCManager.v1_0_0.Manager",
"Certificates": {
"@odata.id": "/redfish/v1/Managers/bmc/Truststore/Certificates"
}
}
},
"UUID": "40575e98-90d7-4c10-9eb5-8d8a7156c9b9"
}
```
Change-Id: Ic82aa5fe760eda31e2792fbdfb6884ac3ea613dc
Signed-off-by: Rohit PAI <rohitpai77@gmail.com>
|
|
Abstracting Node can help us extend Tries for other use case like sub
routes management
Tested
1. Service Validator passes
Change-Id: I4703af9f30107ce2bc3685683a5fd5b669341d35
Signed-off-by: Rohit PAI <ropai@nvidia.com>
|
|
Just as the title says. Trie is useful outside of just the router (for
making other routers.)
Change-Id: I961927f2cea44ee78f32337e64741edad9dc542f
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
This trie class would be useful to use with any arbitrary type, not just
those inheriting from BaseRule.
Change-Id: I325474a100e083ea36407530c6e4e8f6412718ac
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
302 cache handling appears to have been broken when we went to AsyncResp
in the connection class instead of using Response directly. This is
because the expected hash was being written to the old response, not the
new one.
Resolve the issue.
Tested:
using curl to grab /redfish/v1
then pull the ETAG from the response
then use curl to set if-none-match
Shows that redfish now responds with 302 not modified.
Loading a browser window shows many requests are fulfilled with 302 not
modified.
Change-Id: Ie1e782fd5b2c6a5bcf942849ee13ca074973bf1e
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
As the comment states, openbmc currently squashes DEBUG level messages.
After ee993dc84b1e9917b545fdd7367f1127a358084a systemd can see the log
levels, which has the unintended consequence of squashing DEBUG level
messages when enabled.
This is a temporary workaround to fix the regression. Ulimately
implementing something like DEBUG_INVOCATION might be the way to go.
Tested: Enabled debug logging
journalctl -u bmcweb showed debug level messages.
Change-Id: I3c57a47282dbcbf34c58a12d2b7da54f1082fac1
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Generated enum files include `json.hpp' which causes the CI failures
like
```
/var/jenkins-openbmc/workspace/ci-repository/openbmc/bmcweb/redfish-core/include/generated/enums/acceleration_function.hpp:4:1: error: included header json.hpp is not used directly [misc-include-cleaner,-warnings-as-errors]
4 | #include <nlohmann/json.hpp>
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
5 |
```
So, this will treat `<nlohmann/json.hpp>` as `misc-include-cleaner` to
ignore the header check.
Tested:
- Compiles good
- CI passes
Change-Id: Id1d5e981b1de37398aa02c3303357e66fe902efd
Signed-off-by: Myung Bae <myungbae@us.ibm.com>
|
|
This commit is intended to implement the UserPrincipalName (UPN) parse
mode on mutual TLS (MTLS). By implementing this we can use the X509
certificate extension Subject Alternative Name (SAN), specifically UPN
to be used as the username
In our case, this feature is needed because we have a specific format
on our Subject CN of X509 certificate. This format cannot directly
mapped to the username of bmcweb because it contains special
characters (`/` and `:`), which cannot exist in the username.
Changing the format of our Subject CN is very risky. By enabling
this feature we can use other field, which is the SAN extension to
be used as the username and do not change our Subject CN on the
X509 certificate
In general, by implementing this feature, we can enable multiple
options for the system. There might be other cases where we want to
have the username of the bmcweb is not equal to the Subject CN of the
certificate, instead the username is added as the UserPrincipalName
field in the certificate
The format of the UPN is `<username>@<domain>` [1][2]. The format
is similar to email format. The domain name identifies the domain
in which the user is located [3] and it should match the device name's
domain (domain forest).
Tested
- Test using `generate_auth_certificate.py` (extended on patch [4])
- Manual testing (please see the script mentioned above for more detail)
- Setup certificate with UPN inside SAN extension
- Change the CertificateMappingAttribute to use UPN
- Get request to `/SessionService/Sessions`
- Run unit tests
[1] UPN Format: https://learn.microsoft.com/en-us/windows/win32/secauthn/user-name-formats#user-principal-name
[2] UPN Properties: https://learn.microsoft.com/en-us/windows/win32/ad/naming-properties#userprincipalname
[3] UPN Glossary: https://learn.microsoft.com/en-us/openspecs/windows_protocols/ms-wcce/719b890d-62e6-4322-b9b1-1f34d11535b4#gt_9d606f55-b798-4def-bf96-97b878bb92c6
[4] Patch Testing Script: https://gerrit.openbmc.org/c/openbmc/bmcweb/+/78837
Change-Id: I490da8b95aee9579546971e58ab2c4afd64c5997
Signed-off-by: Malik Akbar Hashemi Rafsanjani <malikrafsan@meta.com>
|
|
Verify similar to beb96b0 Break out websockets
Break out the SSE functions into a separate compile unit. This allows
the SSE sockets in beast to be compiled separately, which significantly
reduces the overall compile time by a few seconds. Code is identical
with the exceptions of minor header definitions to convert header-only
to compile unit.
Change-Id: I5aae4f17cbd2badf75b3e0bb644a2309f6300663
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
The commit [1] altered how sockets are created and inadvertently removed
the default port 18080. We use this port extensively in development.
The bmcweb documentation describes the port 18080 for this use. [2]
Adding back the default port 18080.
The commit [1] added meson build options for adding additional ports.
In attempting to enable port 18080 using that mechanism I ran into
various build errors when additional-ports has a value. I've corrected
the config/meson.build file to address those errors. These changes are
not necessary to re-enable port 18080 but worth fixing anyway.
Note: Meson defines arrays as containing strings so there is no
to_string() method for the array member. [3]
```
../../../../../../workspace/sources/bmcweb/config/meson.build:137:39: ERROR: Unknown method "to_string" in object <[StringHolder] holds [str]: '18080'> of type StringHolder.
```
Tested:
- Started bmcweb from /tmp and was able to connect through port 18080.
Log shows:
```
[DEBUG app.hpp:111] Got 0 sockets to open
[INFO app.hpp:150] Starting webserver on port 18080
```
[1] https://gerrit.openbmc.org/c/openbmc/bmcweb/+/35265
[2] https://github.com/openbmc/bmcweb/blob/cc67d0a0fed101c930b334a583d9ca9b222ceb77/TESTING.md?plain=1#L57
[3] https://mesonbuild.com/Build-options.html#arrays
Change-Id: Ia1b326bedca808e43e73ce2b241178bc4bfab23c
Signed-off-by: Janet Adkins <janeta@us.ibm.com>
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Fix minor #include regression caused by
ee993dc84b1e9917b545fdd7367f1127a358084a
Change-Id: Ieda0205a4a4faf877a4f2298e6935bc3aa506fde
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Systemd has an option[1] that allows it to interpret our log levels
directly. This allows for journald to sort/filter/colorize our logs
better than it was able to previously. Its indexes don't map perfectly
to bmcwebs, so come up with a constexpr lookup table to map the two
values across.
Tested: Enabled logging, and dumped journal logs. Observed colorized
output.
[1] https://www.freedesktop.org/software/systemd/man/latest/systemd.exec.html#SyslogLevelPrefix=
Change-Id: I7722ae86e114daec88709b68405498eeb8164c07
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
To support HTTP2 simultaneously on http and https connections, the HTTP
connection classes formerly took the socket as a template option,
allowing passing ssl::stream<tcp::socket> or simply tcp socket. With
the addition of the multiple-sockets option, this would cause two copies
of the template to be instantiated, increasing both compile times and
binary size.
This commit applies the same logic to http2connection as was applied to
HTTPConnection, adding an http type parameter to the constructor, which
allows switching between adapter and adapter.next_level() on each read
or write operation. In compiled code, this means that the connection
classes are only specialized once.
Tested:
When configured for one of each http and https socket and http2
curl --http2 http://<ip>/redfish/v1
succeeds
curl --http2 https://<ip>/redfish/v1 succeeds
Change-Id: I8f33796edd5874d5b93d10a3f253cfadd4f6d7a4
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
When running clang build analyzer it shows that one of the largest
templates in the Redfish compile unit (which is also the slowest compile
unit) is actually the beast websocket instantiations, taking about 6
seconds to compile on their own.
Luckily, the websocket layer is already split into Websocket and
WebsocketImpl classes that we inherited from crow. Unfortunately, crow
puts these two classes into the same file. So.
1. Move the WebSocketImpl class into its own header.
2. Move the websocket() upgrade routes in the websocket routing rule
into their own compile unit to take advantage of only needing
WebsocketImpl in this compile unit.
Tested: Drops build time by several seconds depending on what other
level of optimizations are present (1:15 -> 1:00)
[1] https://github.com/aras-p/ClangBuildAnalyzer
Change-Id: Ia0445eae4a793bb4ccb28136f30d2a05662c529c
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
At one point it was thought that we could pass the io_context object
through the request object, and have the potential to run multiple
io_context instances (one per connection).
Given the safety refactoring we had to do in
9838eb20341568971b9543c2187372d20daf64aa that idea is on ice for the
moment, and would need a major rethink of code to be viable. For the
moment, and in prep for
https://gerrit.openbmc.org/c/openbmc/bmcweb/+/75668
make sure all calls are pulling from the same io object.
Tested: Unit tests pass. Redfish service validator passes.
Change-Id: I877752005c4ce94efbc13ce815f3cd0d99cc3d51
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Using the utility classes reduces compile times and reduces the number
of template specializations that get generated. These are the last two
left in the codebase for getProperty, so fix them.
Tested: On Last commit.
Change-Id: I0ca8411b74b58dbeb42587b88cfd66c0e674b8c8
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
This commit attempts to add the concept of an SSL detector from beast,
and add the capability into bmcweb. This allows directing multiple
socket files to the bmcweb instance, and bmcweb will automatically sort
out whether or not they're SSL, and give the correct response. This
allows users to plug in erroneous urls like "https://mybmc:80" and they
will forward and work correctly.
Some key design points:
The HTTP side of bmcweb implements the exact same http headers as the
HTTPS side, with the exception of HSTS, which is explicitly disallowed.
This is for consistency and security.
The above allows bmcweb builds to "select" the appropriate security
posture (http, https, or both) for a given channel using the
FileDescriptorName field within a socket file. Items ending in:
both: Will support both HTTPS and HTTP redirect to HTTPS
https: Will support HTTPS only
http: will support HTTP only
Given the flexibility in bind statements, this allows administrators to
support essentially any security posture they like. The openbmc
defaults are:
HTTPS + Redirect on both ports 443 and port 80 if http-redirect is
enabled
And HTTPS only if http-redirect is disabled.
This commit adds the following meson options that each take an array of
strings, indexex on the port.
additional-ports
Adds additional ports that bmcweb should listen to. This is always
required when adding new ports.
additional-protocol
Specifies 'http', 'https', or 'both' for whether or not tls is enfoced
on this socket. 'both' allows bmcweb to detect whether a user has
specified tls or not on a given connection and give the correct
response.
additional-bind-to-device
Accepts values that fill the SO_BINDTODEVICE flag in systemd/linux,
and allows binding to a specific device
additional-auth
Accepts values of 'auth' or 'noauth' that determines whether this
socket should apply the normal authentication routines, or treat the
socket as unauthenticated.
Tested:
Previous commits ran the below tests.
Ran the server with options enabled. Tried:
```
curl -vvvv --insecure --user root:0penBmc http://192.168.7.2/redfish/v1/Managers/bmc
* Trying 192.168.7.2:80...
* Connected to 192.168.7.2 (192.168.7.2) port 80 (#0)
* Server auth using Basic with user 'root'
> GET /redfish/v1/Managers/bmc HTTP/1.1
> Host: 192.168.7.2
> Authorization: Basic cm9vdDowcGVuQm1j
> User-Agent: curl/7.72.0
> Accept: */*
>
* Mark bundle as not supporting multiuse
< HTTP/1.1 301 Moved Permanently
< Location: https://192.168.7.2
< X-Frame-Options: DENY
< Pragma: no-cache
< Cache-Control: no-Store,no-Cache
< X-XSS-Protection: 1; mode=block
< X-Content-Type-Options: nosniff
< Content-Security-Policy: default-src 'none'; img-src 'self' data:; font-src 'self'; style-src 'self'; script-src 'self'; connect-src 'self' wss:
< Date: Fri, 08 Jan 2021 01:43:49 GMT
< Connection: close
< Content-Length: 0
<
* Closing connection 0
```
Observe above:
webserver returned 301 redirect.
webserver returned the appropriate security headers
webserver immediately closed the connection.
The same test above over https:// returns the values as expected
Loaded the webui to test static file hosting. Webui logs in and works
as expected.
Used the scripts/websocket_test.py to verify that websockets work.
Sensors report as expected.
Change-Id: Ib5733bbe5473fed6e0e27c56cdead0bffedf2993
Signed-off-by: Ed Tanous <ed@tanous.net>
|
|
h2c upgrade is a mechanism for supporting http/2 on connections that
might not support alpn [1]. This is done by the client specifying
Connection: upgrade
Upgrade: h2c
This looks very similar to a websocket upgrade, which h2c replacing
websocket. Because of this, the existing upgrade code needs some
upgrades to avoid parsing twice.
Tested:
```
curl -u root:0penBmc --http2 -k http://192.168.7.2:443/redfish/v1/SessionService/Sessions
```
Succeeds and verbose logging shows that http upgrade succeeded
websocket_test.py in the scripts directory connects and reports events
[1] https://datatracker.ietf.org/doc/html/rfc7540#section-11.8
Change-Id: I8f76e355f99f21337d310ef2f345e6aaa253b48b
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
base64 decoding comes in two flavors, "normal" which we already
implement, and "url safe" which modifies the alphabet to create base64
encodings that are safe to use in filenames and urls. Functionally this
just involves swapping two characters with underscore and minus in the
encode/decode table. To avoid duplicating a lot of code, this commit
refactors the base64 tables to be generated at compile time.
Tested: Included unit tests pass. No usage until next commit.
Change-Id: I71724fd2e04000f115c22a40d382d411986d7b39
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Copy the latest format file from the docs repository and apply.
Change-Id: I2f0b9d0fb6e01ed36a2f34c750ba52de3b6d15d1
Signed-off-by: Patrick Williams <patrick@stwcx.xyz>
|
|
The way we pass around io contexts is somewhat odd. Boost maintainers
in slack recommended that we just have a method that returns an io
context, and from there we can control this (context link lost years
ago).
The new version of clang claims the singleton pattern of passing in an
io_context pattern is a potential nullptr dereference. It's technically
correct, as calling the singleton without immediately initializing the
io context will lead to a crash.
This commit implements what the boost maintainers suggested, having a
single method that returns "the context" that should be used. This also
helps to maintain isolation, as some pieces are no longer tied directly
to dbus to get their reactor.
Tested: WIP
Change-Id: Ifaa11335ae00a3d092ecfdfb26a38380227e8576
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Clang-tidy misc-include-cleaner appears to now be enforcing
significantly more headers than previously. That is overall a good
thing, but forces us to fix some issues. This commit is largely just
taking the clang-recommended fixes and checking them in. Subsequent
patches will fix the more unique issues.
Note, that a number of new ignores are added into the .clang-tidy file.
These can be cleaned up over time as they're understood. The majority
are places where boost includes a impl/x.hpp and x.hpp, but expects you
to use the later. include-cleaner opts for the impl, but it isn't clear
why.
Change-Id: Id3fdd7ee6df6c33b2fd35626898523048dd51bfb
Signed-off-by: Ed Tanous <etanous@nvidia.com>
Signed-off-by: Gunnar Mills <gmills@us.ibm.com>
|
|
SPDX identifiers are simpler, and reduce the amount of cruft we have in
code files. They are recommended by linux foundation, and therefore we
should do as they allow.
This patchset does not intend to modify any intent on any existing
copyrights or licenses, only to standardize their inclusion.
[1] https://www.linuxfoundation.org/blog/blog/copyright-notices-in-open-source-software-projects
Change-Id: I935c7c0156caa78fc368c929cebd0f068031e830
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
These two parameters are not idempotent, and are currently being run
first on the satellite, then again on the aggregator. This results in
errors and unexpected results.
This commit detects when we're parsing a top collection, and filters out
those two parameters from being applied to the satellite request.
To accomplish rewriting the URI, a new Request API needs added for
non-const access to the URI object.
Tested: Aggregator shows results as expected. Query params are not
forwarded to satellite
Change-Id: I99cbbb08da9fcd06c9ee10d371b253e32d01f59b
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Update to add new checks that are now available to us. Fix the minor
issues we have. A few of our checks that we previously had enabled have
been renamed, so remove those from the file as well.
Change-Id: Idbbfc3cb7ba42ac780e557554d7ae8ab190e7551
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Previously this function was based on a basic string comparison. This
is fine, but found several inconsistencies, like not handling spaces in
the appropriate places.
This commit creates a new function getContentType, using the new parsing
infrastructure. As doing this, it showed that the existing parser
functions were not handling case insensitive compares for the mime type.
While this is technically not required, it's something we unit test for,
and relatively easy to add.
Note, that because this parser ignores charset, this moves charset=ascii
from something that previously failed, to something that now succeeds.
This is expected.
Tested: Unit tests pass. Good coverage
Change-Id: I825a72862135b62112ee504ab0d9ead9d6796354
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Coverity marks some minor things as improvements we can make.
Clean up the code to silence the errors.
Tested: Unit tests coverage for http core is sufficient.
Change-Id: If4efb359792bfdfe3866e843b4bbdb3f83fec0c5
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
When sending last-event-id, the previous events were being received
before the the header was completed. This is because the open handler is
being called before the connect call was in place, so you get:
Connection starts
open handler called
sendEvent() called from open handler
sendSSEHeader() called.
This results in a spec violation.
Tested:
curl --user root:0penBmc -vvv -s --no-buffer -k -N -H 'Accept:
text/event-stream' -H 'Last-Event-Id: 4' -X GET
https://localhost:8000/redfish/v1/EventService/SSE
Now succeeds, and wireshark dumps show the header is being sent
correctly. Note that previously this command would fail unless http0.9
header was set.
Unit test coverage for this path without last-event-id passes.
Change-Id: I44bb6eedbcbdc727b257646ec55e808157231f75
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Change-Id: Iefe1b695b86a640d8dfaafd1f77f374fa34246de
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
Change-Id: I6d677b16219482db16c64d5d8412ca557142a597
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
|
As reported, there are cases where a valid certificate isn't present,
but a browser still prompts for an MTLS cert. Fix that by explicitly
setting verify_none if strict tls isn't enabled. Unclear what impacts
this will have elsewhere:
Tested (not yet done on this patch): with a self-signed certificate,
logging into chrome no longer prompts the certificate screen.
Change-Id: Iaf7d25fec15ad547a6c741c9410995e19ba22016
Signed-off-by: Ed Tanous <etanous@nvidia.com>
|
