- 14 Jan, 2020 17 commits
-
-
Guillaume Nault authored
When peernet2id() had to lock "nsid_lock" before iterating through the nsid table, we had to disable BHs, because VXLAN can call peernet2id() from the xmit path: vxlan_xmit() -> vxlan_fdb_miss() -> vxlan_fdb_notify() -> __vxlan_fdb_notify() -> vxlan_fdb_info() -> peernet2id(). Now that peernet2id() uses RCU protection, "nsid_lock" isn't used in BH context anymore. Therefore, we can safely use plain spin_lock()/spin_unlock() and let BHs run when holding "nsid_lock". Signed-off-by:Guillaume Nault <gnault@redhat.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Guillaume Nault authored
__peernet2id() can be protected by RCU as it only calls idr_for_each(), which is RCU-safe, and never modifies the nsid table. rtnl_net_dumpid() can also do lockless lookups. It does two nested idr_for_each() calls on nsid tables (one direct call and one indirect call because of rtnl_net_dumpid_one() calling __peernet2id()). The netnsid tables are never updated. Therefore it is safe to not take the nsid_lock and run within an RCU-critical section instead. Signed-off-by:
-
Guillaume Nault authored
__peernet2id_alloc() was used for both plain lookups and for netns ID allocations (depending the value of '*alloc'). Let's separate lookups from allocations instead. That is, integrate the lookup code into __peernet2id() and make peernet2id_alloc() responsible for allocating new netns IDs when necessary. This makes it clear that __peernet2id() doesn't modify the idr and prepares the code for lockless lookups. Also, mark the 'net' argument of __peernet2id() as 'const', since we're modifying this line. Signed-off-by:
-
Geert Uytterhoeven authored
Convert mdiobus_register_reset() from open-coded DT-only optional reset handling to reset_control_get_optional_exclusive(). This not only simplifies the code, but also adds support for lookup-based resets on non-DT systems. Signed-off-by:
Geert Uytterhoeven <geert+renesas@glider.be> Reviewed-by:
Andrew Lunn <andrew@lunn.ch> Reviewed-by:
Philipp Zabel <p.zabel@pengutronix.de> Signed-off-by:
-
Florian Fainelli authored
The information about the PHY attached to the PHYLINK instance is useful but is missing the IRQ prints that phy_attached_info() adds. phy_attached_info() is a bit long and it would not be possible to use phylink_info() anyway. Signed-off-by:
Florian Fainelli <f.fainelli@gmail.com> Signed-off-by:
-
Jakub Kicinski authored
Jose Abreu says: ==================== net: stmmac: ETF support This series adds the support for ETF scheduler in stmmac. 1) Starts adding the support by implementing Enhanced Descriptors in stmmac main core. This is needed for ETF feature in XGMAC and QoS cores. 2) Integrates the ETF logic into stmmac TC core. 3) and 4) adds the HW specific support for ETF in XGMAC and QoS cores. The IP feature is called TBS (Time Based Scheduling). 5) Enables ETF in GMAC5 IPK PCI entry for all Queues except Queue 0. 6) Adds the new TBS feature and even more information into the debugFS HW features file. ==================== Signed-off-by:Jakub Kicinski <kuba@kernel.org>
-
Jose Abreu authored
Add a new test for TBS feature which is used in ETF scheduler. In this test, we send a packet with a launch time specified as now + 500ms and check if the packet was transmitted on that time frame. Changes from v2: - Use the TBS bitfield - Remove debug message Signed-off-by:
Jose Abreu <Jose.Abreu@synopsys.com> Signed-off-by:
-
Jose Abreu authored
In the upcoming commit for TBS selftest we will need to send a packet on a specific Queue. As stmmac fallsback to netdev_pick_tx() on the select Queue callback, we need to switch all selftests logic to dev_direct_xmit() so that we can send the given SKB on a specific Queue. Signed-off-by:
-
Jose Abreu authored
Adds more information regarding HW Capabilities in the corresponding DebugFS file. Changes from v2: - Remove the TX/RX queues in use (Jakub) Signed-off-by:
-
Jose Abreu authored
Enable TBS support on GMAC5 PCI entry for all Queues except Queue 0. Signed-off-by:
-
Jose Abreu authored
Adds all the necessary HW hooks to support TBS feature in QoS cores. Changes from v1: - Remove unneeded LT shift as the IP already does this. Signed-off-by:
-
Jose Abreu authored
Adds all the necessary HW hooks to support TBS feature in XGMAC cores. Changes from v1: - Remove unneeded LT shift as the IP already does this. Signed-off-by:
-
Jose Abreu authored
Adds the support for ETF scheduler using TBS feature which is available in XGMAC and QoS IPs. Changes from v2: - Fix checkpatch issues (Jakub) - Use the TBS bitfield Signed-off-by:
-
Jose Abreu authored
Adds the initial hooks for TBS support. This needs a 32 byte descriptor in order for it to work with current HW. Adds all the logic for Enhanced Descriptors in main core but no HW related logic for now. Changes from v2: - Use bitfield for TBS status / support (Jakub) - Remove unneeded cache alignment (Jakub) - Fix checkpatch issues Signed-off-by:
-
Chen Zhou authored
The conversion to bool is not needed, remove it. Signed-off-by:
Chen Zhou <chenzhou10@huawei.com> Signed-off-by:
-
Jesper Dangaard Brouer authored
Commit 0bf7800f ("ptr_ring: try vmalloc() when kmalloc() fails") started to use kvmalloc_array and kvfree, which are defined in mm.h, the previous functions kcalloc and kfree, which are defined in slab.h. Add the missing include of linux/mm.h. This went unnoticed as other include files happened to include mm.h. Fixes: 0bf7800f ("ptr_ring: try vmalloc() when kmalloc() fails") Signed-off-by:
Jesper Dangaard Brouer <brouer@redhat.com> Acked-by:
Michael S. Tsirkin <mst@redhat.com> Signed-off-by:
-
Lorenzo Bianconi authored
With 'commit 44768dec ("page_pool: handle page recycle for NUMA_NO_NODE condition")' we can safely change nid to NUMA_NO_NODE and accommodate future NUMA aware hardware using mvneta network interface Signed-off-by:
Lorenzo Bianconi <lorenzo@kernel.org> Acked-by:
-
- 13 Jan, 2020 3 commits
-
-
Zhang Xiaoxu authored
Fix sparse warning: drivers/net/ethernet/sfc/ethtool_common.c warning: symbol 'efx_fill_test' was not declared. Should it be static? warning: symbol 'efx_fill_loopback_test' was not declared. Should it be static? warning: symbol 'efx_describe_per_queue_stats' was not declared. Should it be static? Reported-by:Hulk Robot <hulkci@huawei.com> Signed-off-by:
Zhang Xiaoxu <zhangxiaoxu5@huawei.com> Reviewed-by:
Martin Habets <mhabets@solarflare.com> Signed-off-by:
-
Geert Uytterhoeven authored
Use the print_hex_dump_debug() helper, instead of open-coding the same operations. Signed-off-by:
Geert Uytterhoeven <geert@linux-m68k.org> Signed-off-by:
-
Geert Uytterhoeven authored
Remove unused fields, copied from the Sun LANCE driver eons ago. Signed-off-by:
-
- 12 Jan, 2020 11 commits
-
-
Jakub Kicinski authored
Linus Walleij says: ==================== IXP4xx networking cleanups This is a patch series which jams together Arnds and mine cleanups for the IXP4xx networking. I also have patches for device tree support but that requires more elaborate work, this series is some of mine and some of Arnds patches that is a good foundation for his multiplatform work and my device tree work. These are for application to the networking tree so that can be taken in one separate sweep. I have tested the patches for a bit using zeroday builds and some boots on misc IXP4xx devices and haven't run into any major problems. We might find some new stuff as a result from the new compiler coverage. I had to depromote enabling compiler coverage at one point in the v2 set because it depended on other patches making the code more generic. The change in v3 was simply dropping one offending patch hardcoding base addresses into the driver. The change in v4 drops a stable@ tag that was unnecessary. This v5 is a rebase of the v4 patch set on top of net-next. ==================== Signed-off-by: -
Linus Walleij authored
Use the netdevice struct device .parent field when calling dma_pool_create(): the .dma_coherent_mask and .dma_mask pertains to the bus device on the hardware (platform) bus in this case, not the struct device inside the network device. This makes the pool allocation work. Signed-off-by:
Linus Walleij <linus.walleij@linaro.org> Signed-off-by:
-
Linus Walleij authored
In order to probe this ethernet interface from the device tree all physical MMIO regions must be passed as resources. Begin this rewrite by first passing the port base address as a resource for all platforms using this driver, remap it in the driver and avoid using any reference of the statically mapped virtual address in the driver. Signed-off-by:
-
Linus Walleij authored
Simplify and correct a bunch of messages using printk directly to use the netdev_* macros. I have not changed all of them, just the low-hanging fruit. Signed-off-by:
-
Linus Walleij authored
Use "ndev" for the struct net_device and "dev" for the struct device in probe() and remove(). Add the local "dev" pointer for later use in refactoring. Take this opportunity to fix inverse christmas tree coding style. Signed-off-by:
-
Linus Walleij authored
The IXP4xx driver was initializing the MDIO bus before even probing, in the callbacks supposed to be used for setting up the module itself, and with the side effect of trying to register the MDIO bus as soon as this module was loaded or compiled into the kernel whether the device was discovered or not. This does not work with multiplatform environments. To get rid of this: set up the MDIO bus from the probe() callback and remove it in the remove() callback. Rename the probe() and remove() calls to reflect the most common conventions. Since there is a bit of checking for the ethernet feature to be present in the MDIO registering function, making the whole module not even be registered if we can't find an MDIO bus, we need something similar: register the MDIO bus when the corresponding ethernet is probed, and return -EPROBE_DEFER on the other interfaces until this happens. If no MDIO bus is present on any of the registered interfaces we will eventually bail out. None of the platforms I've seen has e.g. MDIO on EthB and only uses EthC, there is always a Ethernet hardware on the NPE (B, C) that has the MDIO bus, we just might have to wait for it. Signed-off-by:
-
Arnd Bergmann authored
The platform data is needed to compile the driver as standalone, so move it to a global location along with similar files. Signed-off-by:
Arnd Bergmann <arnd@arndb.de> Signed-off-by:
-
Arnd Bergmann authored
The ixp46x ptp driver has a somewhat unusual setup, where the ptp driver and the ethernet driver are in different directories but access the same registers that are defined a platform specific header file. Moving everything into drivers/net/ makes it look more like most other ptp drivers and allows compile-testing this driver on other targets. Signed-off-by:
-
Arnd Bergmann authored
The ixp4xx_hss driver needs the platform data definition and the system clock rate to be compiled. Move both into a new platform_data header file. This is a prerequisite for compile testing, but turning on compile testing requires further patches to isolate the SoC headers. Signed-off-by:
-
Arnd Bergmann authored
Change the driver to use portable integer types to avoid warnings during compile testing: drivers/net/wan/ixp4xx_hss.c:863:21: error: cast to 'u32 *' (aka 'unsigned int *') from smaller integer type 'int' [-Werror,-Wint-to-pointer-cast] memcpy_swab32(mem, (u32 *)((int)skb->data & ~3), bytes / 4); ^ drivers/net/wan/ixp4xx_hss.c:979:12: error: incompatible pointer types passing 'u32 *' (aka 'unsigned int *') to parameter of type 'dma_addr_t *' (aka 'unsigned long long *') [-Werror,-Wincompatible-pointer-types] &port->desc_tab_phys))) ^~~~~~~~~~~~~~~~~~~~ include/linux/dmapool.h:27:20: note: passing argument to parameter 'handle' here dma_addr_t *handle); ^ Signed-off-by: -
Jonathan Lemon authored
On modern hardware with a large number of cpus and using XDP, the current MSIX limit is insufficient. Bump the limit in order to allow more queues. Signed-off-by:
Jonathan Lemon <jonathan.lemon@gmail.com> Reviewed-by:
Jack Wang <jinpu.wang@cloud.ionos.com> Reviewed-by:
Tariq Toukan <tariqt@mellanox.com> Signed-off-by:
-
- 11 Jan, 2020 9 commits
-
-
David S. Miller authored
Huazhong Tan says: ==================== net: hns3: add some misc update about reset issue This series includes some misc update relating to reset issue. [patch 1/7] & [patch 2/7] splits hclge_reset()/hclgevf_reset() into two parts: preparing and rebuilding. Since the procedure of FLR should be separated out from the reset task([patch 3/7 & patch 3/7]), then the FLR's processing can reuse these codes. pci_error_handlers.reset_prepare() is void type function, so [patch 6/7] & [patch 7/7] factor some codes related to PF function reset to make the preparing done before .reset_prepare() return. BTW, [patch 5/7] enlarges the waiting time of reset for matching the hardware's. ==================== Signed-off-by: -
Huazhong Tan authored
hclge_reset_prepare_down() is only used to inform VF that PF is going to do function reset, then using hclge_func_reset_sync_vf() in hclge_reset_prepare_wait() to query whether VF is ready before asserting PF function reset. To make the code more readable, this patch uses a new function hclge_function_reset_notify_vf() to do this job. Signed-off-by:
Huazhong Tan <tanhuazhong@huawei.com> Signed-off-by:
-
Huazhong Tan authored
When synchronizes with VFs fail before PF function reset, PF driver should go on its function reset, otherwise it can not run normally anymore. So, hclge_func_reset_sync_vf() should not affect the processing of PF reset, this patch modifies its return type to void. Signed-off-by:
-
Huazhong Tan authored
When the load of firmware is high, its reset task may takes more time(which will be as long as 35 seconds). So this patch modifies HCLGE_RESET_WAIT_CNT to match the firmware's. Signed-off-by:
-
Huazhong Tan authored
Currently, the actual work of VF FLR is handled in the reset task, which is asynchronous. So in some case, if the preparing and rebuilding are not done, then the VF FLR will trigger some problems, for example, makes hardware go into chaos. So this patch separates the process of VF FLR from reset task, and adds a semaphore to serialize this reset and others. When FLR's preparing fails, if there has other higher level reset pending or failing times less than the HCLGE_FLR_RETRY_CNT, this preparing should be retried, otherwise it will get into a wrong state. BTW, while the hardware reports misc interrupt during pcie_flr(), the driver can not receive this interrupt anymore, so disable it when hclgevf_flr_prepare() return, and re-enable it when enter hclgevf_flr_done(). Avoid declaring internal function hclgevf_enable_vector(), this patch also moves its definition forward, and removes unused enum hnae3_flr_state. Signed-off-by:
-
Huazhong Tan authored
Currently, the actual work of PF FLR is handled in the reset task, which is asynchronous. So in some case, if the preparing and rebuilding are not done, then the PF FLR will trigger some problems, for example, makes hardware go into chaos. So this patch separates the process of PF FLR from reset task, and adds a semaphore to serialize this reset and others. When FLR's preparing fails, if there has other higher level reset pending or failing times less than the HCLGE_FLR_RETRY_CNT, this preparing should be retried, otherwise PF and its VF may get into wrong state. BTW, while the hardware reports misc interrupt during pcie_flr(), the driver can not receive this interrupt anymore, so disable it when hclge_flr_prepare() return, and re-enable it when enter hclge_flr_done(). Signed-off-by:
-
Huazhong Tan authored
hclgevf_reset() is a little bloated, and the process of VF FLR will be separated from the reset task later. So this patch splits hclgevf_reset() into hclgevf_reset_prepare() and hclge_reset_rebuild(), then FLR can also reuse these two functions. Also moves HNAE3_UP_CLIENT into hclgevf_reset_stack(). Signed-off-by:
-
Huazhong Tan authored
hclge_reset() is a little bloated, and the process of PF FLR will be separated from the reset task later. So this patch splits hclge_reset() into hclge_reset_prepare() and hclge_reset_rebuild(), then FLR can also reuse these two functions. BTW, since hclge_clear_reset_cause() and hclge_reset_prepare_up() will not affect the device, so in hclge_reset_rebuild(), these functions are called without rtnl_lock. Signed-off-by:
-
YueHaibing authored
Fixes gcc '-Wunused-but-set-variable' warning: drivers/net/ethernet/sfc/mcdi_functions.c: In function 'efx_mcdi_ev_init': drivers/net/ethernet/sfc/mcdi_functions.c:79:28: warning: variable 'nic_data' set but not used [-Wunused-but-set-variable] commit 4438b587 ("sfc: move MCDI event queue management code") introduces this unused variable. Reported-by:
YueHaibing <yuehaibing@huawei.com> Signed-off-by:
-
