- 01 Nov, 2016 20 commits
-
-
Jakub Kicinski authored
We used to always allocate the same number of TX and RX rings so the support for having r_vectors without one of the rings was dropped. That makes us, however, unnecessarily limited to 8 TX rings (8 is the Linux RSS default) most of the time. Also we are about to add channel count configuration via ethtool, so bring that support back. TX rings can now default to num_online_cpus() and RX rings to netif_get_num_default_rss_queues(). Signed-off-by:
Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Jakub Kicinski authored
num_irqs is not used anywhere, replace it with max_r_vecs which holds number of allocated RX/TX vectors and is going to be useful soon. Signed-off-by:
-
Jakub Kicinski authored
nfp_net_irqs_wanted() doesn't really encapsulate much logic, remove it and inline the calculations. Signed-off-by:
-
Jakub Kicinski authored
Use unsigned int consistently for vector/ring counts. Signed-off-by:
-
Jakub Kicinski authored
We are currently using define for max TX rings to allocate IRQ vectors. It's OK since the max number of rings for TX and RX are currently the same, but lets make the code nicer by taking max of the two. Signed-off-by:
-
Jakub Kicinski authored
We already force ring sizes to be power of 2 so replace modulo operations with AND (size - 1) in index calculations. Signed-off-by:
-
Jakub Kicinski authored
Introduce a separate buffer allocation function to be called from NAPI. We can make assumptions about the context and buffer size. Signed-off-by:
-
Jakub Kicinski authored
Speed up RX processing by moving to the alloc_frag()/build_skb() paradigm. Since we're no longer mapping the entire buffer for DMA add helpers which take care of calculating offsets and lengths. Signed-off-by:
-
Jakub Kicinski authored
nfp_net_rx() is quite long already and about to get longer. Move buffer drop/recycle to a helper. Signed-off-by:
-
Jakub Kicinski authored
Add a helper function to calculate the buffer size at run time. Buffer lengths will now depend on the FW prepend configuration instead of assuming the most space consuming configuration and defaulting to 2k buffers at initialization time. Signed-off-by:
-
Jakub Kicinski authored
Don't declare functions as static inline in .c files and remove dead code it was hiding. Signed-off-by:
-
Jakub Kicinski authored
ether_setup() will be invoked by alloc_etherdev_mqs(), no need to call it again. Signed-off-by:
-
Jakub Kicinski authored
Drop all code related to nfp3200. It was never widely deployed as a NIC. Signed-off-by:
-
Jakub Kicinski authored
There are few variables in nfp_net_poll() which are used only once or unused but set. Remove them. Signed-off-by:
-
David S. Miller authored
Yuval Mintz says: ==================== qed*: Fixes to "Allow unicast filtering" Commit 7b7e70f9 ("qed*: Allow unicast filtering") introduced several issues in driver. This series is intended to address and fix those. ==================== Signed-off-by:
-
Mintz, Yuval authored
When relaxing the limitation on the number of unicast MAC filters an interface can configure, qed started passing the MAC quota to qede. However, the value is initialized only for PFs, causing VFs to always try and configure themselves as promiscuous [as they believe they lack the resources to configure the rx-mode]. Fixes: 7b7e70f9 ("qed*: Allow unicast filtering") Signed-off-by:
Yuval Mintz <Yuval.Mintz@cavium.com> Signed-off-by:
-
Mintz, Yuval authored
Driver is now setting the ndev's priv_flags instead of adding to it, causing pktgen failure to utilize various features due to the loss of the IFF_TX_SKB_SHARING indication. Fixes: 7b7e70f9 ("qed*: Allow unicast filtering") Signed-off-by:
-
git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next-queueDavid S. Miller authored
Jeff Kirsher says: ==================== 40GbE Intel Wired LAN Driver Updates 2016-10-31 This series contains updates to i40e and i40evf. Colin Ian King fixes a minor issue with dev_err message where a new line character was missing from the end of the message. Jake provides several most of the changes in the series, starting with dropping the is_vf and is_netdev fields in the i40e_mac_filter structure since they are not needed (along with the checks that used these fields). Reason being that we use separate VSI's for SRIOV VFs and for netdev VSIs, therefore a single VSI should only have one type of filter. Then simplifies our .set_rx_mode handler by using the kernel provided __dev_uc_sync and __dev_mc_sync functions for notification of add and deletion of filters. Refactored the i40e_put_mac_in_vlan() to resolve an issue where this function was arbitrarily modifying all filters to have the same VLAN, which is incorrect because it could be modifying active filters without putting them into the new state. Refactored the delete filter logic so that we can re-use the functionality, where appropriate, without having to search for the filter twice. Reduced the latency of operations related to searching for specific MAC filters by using a static hash table instead of a list. Reduced code duplication in the adminq command to add/delete for filters. Fixed an issue where TSYNVALID bit was not being checked as the true indicator of whether the packet has an associated timestamp. Cleaned up a second msleep() call by simply re-ordering the code so that the extra wait is no longer needed. Alan provides additional fix to the work Jake has been doing to resolve a bug where adding at least one VLAN and then removing all VLANs leaves the MAC filters for the VSI with an incorrect value for the VID which indicates the MAC filter's VLAN status. Alex adds a common method for finding a VSI by type. Also cleaned up the logic for coalescing RS bits, which was convoluted and larger than it needed to be. Mitch fixes an issue with the failure to add filters when the VF driver is reloaded by simply setting the number of filters to 0 when freeing VF resources. Maciej implements a I40E_NVMUPD_STATE_ERROR state for NVM update, so that the driver has the ability to return NVM image write failure. Filip removes unreachable code which was found using static analysis where "if" statements were never in a "true/false" state, so clean up unnecessary if statements. ==================== Signed-off-by: -
Eric Dumazet authored
Systems with large pages (64KB pages for example) do not always have huge quantity of memory. A big SK_MEM_QUANTUM value leads to fewer interactions with the global counters (like tcp_memory_allocated) but might trigger memory pressure much faster, giving suboptimal TCP performance since windows are lowered to ridiculous values. Note that sysctl_mem units being in pages and in ABI, we also need to change sk_prot_mem_limits() accordingly. Signed-off-by:
Eric Dumazet <edumazet@google.com> Signed-off-by:
-
Paul Gortmaker authored
The Makefile/Kconfig currently controlling compilation of this code is: drivers/net/cris/Makefile:obj-$(CONFIG_ETRAX_ARCH_V10) += eth_v10.o arch/cris/Kconfig:config ETRAX_ARCH_V10 arch/cris/Kconfig: bool ...meaning that it currently is not being built as a module by anyone. Lets remove the couple traces of modular infrastructure use, so that when reading the driver there is no doubt it is builtin-only. Since module_init translates to device_initcall in the non-modular case, the init ordering remains unchanged with this commit. There was a one line wrapper for the int init function, which made no sense; hence we just put the device_initcall on the true init function itself and delete the pointless wrapper. In doing that we get rid of the following compile warning: WARNING: drivers/net/built-in.o(.text+0x1e28): Section mismatch in reference from the function etrax_init_module() to the function .init.text:etrax_ethernet_init() We don't replace module.h with init.h since the file already has that. Cc: "David S. Miller" <davem@davemloft.net> Cc: Mikael Starvik <starvik@axis.com> Cc: Jesper Nilsson <jesper.nilsson@axis.com> Cc: netdev@vger.kernel.org Cc: linux-cris-kernel@axis.com Signed-off-by:
Paul Gortmaker <paul.gortmaker@windriver.com> Signed-off-by:
-
- 31 Oct, 2016 20 commits
-
-
Filip Sadowski authored
Removed some of unnecessary if statements and unreachable code found by static code analysis tool. The return value of i40e_vsi_control_rings(..., false) is always 0. So, test for non-zero will never be true. The function has been split into "int i40e_vsi_start_rings()" and "void i40e_vsi_stop_rings()" for better understanding. Similarly, the function i40e_vsi_kill_vlan() never fails. So, checking for return value is also unnecessary. Function definition changed to void. The i40e_loopback_test() function is not implemented. The function and all references to loopback testing were removed. Change-ID: Id45cf66f6689ce2bc4e887de13f073e30e8431bd Signed-off-by:
Filip Sadowski <filip.sadowski@intel.com> Tested-by:
Andrew Bowers <andrewx.bowers@intel.com> Signed-off-by:
Jeff Kirsher <jeffrey.t.kirsher@intel.com>
-
Maciej Sosin authored
This patch adds I40E_NVMUPD_STATE_ERROR state for NVM update. Without this patch driver has no possibility to return NVM image write failure.This state is being set when ARQ rises error. arq_last_status is also updated every time when ARQ event comes, not only on error cases. Change-ID: I67ce43ba22a240773c2821b436e96054db0b7c81 Signed-off-by:
Maciej Sosin <maciej.sosin@intel.com> Tested-by:
-
Michal Kosiarz authored
For some cases when reading from device are incorrect or image is incorrect, this part of code causes crash due to division by zero. Change-ID: I8961029a7a87b0a479995823ef8fcbf6471405e1 Signed-off-by:
Michal Kosiarz <michal.kosiarz@intel.com> Tested-by:
-
Mitch Williams authored
When a VF is reset, it gets a new VSI, so all of its MAC filters go away. Correctly set the number of filters to 0 when freeing VF resources. This corrects a problem with failure to add filters when the VF driver is reloaded. Change-ID: I2acbecf734287b67473bb225293e14b5096acbef Signed-off-by:
Mitch Williams <mitch.a.williams@intel.com> Tested-by:
-
Alexander Duyck authored
This patch reorders the logic at the end of i40e_tx_map to address the fact that the logic was rather convoluted and much larger than it needed to be. In order to try and coalesce the code paths I have updated some of the comments and repurposed some of the variables in order to reduce unnecessary overhead. This patch does the following: 1. Quit tracking skb->xmit_more with a flag, just max out packet_stride 2. Drop tail_bump and do_rs and instead just use desc_count and td_cmd 3. Pull comments from ixgbe that make need for wmb() more explicit. Change-ID: Ic7da85ec75043c634e87fef958109789bcc6317c Signed-off-by:
Alexander Duyck <alexander.h.duyck@intel.com> Tested-by:
-
Alexander Duyck authored
This patch adds a common method for finding a VSI by type. The main motivation for doing this is that the Flow Director path actually had two ways of handling this, one stopped on first match and one did not. This patch makes it so that all callers of this function will get the same approach for finding a VSI. Change-ID: Ibf25de8acd8466582520694424aa87da66965fbd Signed-off-by:
Bimmy Pujari <bimmy.pujari@intel.com> Tested-by:
-
Jacob Keller authored
Remove the second call to msleep outside the loop, and move the msleep within the loop as the first step. This guarantees that a single loop will wait the minimum time first, and then after the reset finishes we no longer need an extra msleep. Change-ID: Ib2086f0a142402b614f67846bc091754203a0b9a Signed-off-by:
Jacob Keller <jacob.e.keller@intel.com> Tested-by:
-
Jacob Keller authored
The current Rx timestamp hang logic is not very robust because it does not notice a register is hung until all four timestamps have been latched and we wait a full 5 seconds. Replace this logic with a newer Rx hang detection based on storing the jiffies when we first notice a receive timestamp event. We store each register's time separately, along with a flag indicating if it is currently latched. Upon first transitioning to latch, we will update the latch_events[i] jiffies value. This indicates the time we first noticed this event. The watchdog routine will simply check that the either the flag has been cleared, or we have passed at least one second. In this case, it is able to clear the Rx timestamp register under the assumption that it was for a dropped frame. The benefit if this strategy is that we should be able to detect and clear out stalled RXTIME_H registers before we exhaust the supply of 4, and avoid complete stall of Rx timestamp events. Change-ID: Id55458c0cd7a5dd0c951ff2b8ac0b2509364131f Signed-off-by:
-
Jacob Keller authored
We need a locking mechanism to protect the hardware SYSTIME register which is split over 2 values, and has internal hardware latching. We can't allow multiple accesses at the same time. However.... The spinlock_t is overkill here, especially use of spin_lock_irqsave, since every PTP access will halt hardirqs. Notice that the only places which need the SYSTIME value are user context and are capable of sleeping. Thus, it is safe to use a mutex here instead of the spinlock. Change-ID: I971761a89b58c6aad953590162e85a327fbba232 Signed-off-by:
-
Jacob Keller authored
When hardware has taken a timestamp for a received packet, it indicates which RXTIME register the timestamp was placed in by some bits in the receive descriptor. It uses 3 bits, one to indicate if the descriptor index is valid (ie: there was a timestamp) and 2 bits to indicate which of the 4 registers to read. However, the driver currently does not check the TSYNVALID bit and only checks the index. It assumes a zero index means no timestamp, and a non zero index means a timestamp occurred. While this appears to be true, it prevents ever reading a timestamp in RXTIME[0], and causes the first timestamp the device captures to be ignored. Fix this by using the TSYNVALID bit correctly as the true indicator of whether the packet has an associated timestamp. Also rename the variable rsyn to tsyn as this is more descriptive and matches the register names. Change-ID: I4437e8f3a3df2c2ddb458b0fb61420f3dafc4c12 Signed-off-by:
-
Jacob Keller authored
We duplicate some code around adding and deleting filters using the adminq interface. This is prone to errors in case there are bugs. Use functions which extract the logic to their own portion so that we don't duplicate it twice in code. Change-ID: I60d68aeb887976787dec00b23ab386a106e61465 Signed-off-by:
-
Jacob Keller authored
We determine that a VSI is in vlan_mode whenever it has any filters with a VLAN other than -1 (I40E_VLAN_ALL). The previous method of doing so was to perform a loop whenever we needed the check. However, we can notice that only place where filters are added (i40e_add_filter) can change the condition from false to true, and the only place we can return to false is in i40e_vsi_sync_filters_subtask. Thus, we can remove the loop and use a boolean directly. Doing this avoids looping over filters repeatedly especially while we're already inside a loop over all the filters. This should reduce the latency of filter operations throughout the driver. Change-ID: Iafde08df588da2a2ea666997d05e11fad8edc338 Signed-off-by:
-
Alan Brady authored
Currently there exists a bug where adding at least one VLAN and then removing all VLANs leaves the mac filters for the VSI with an incorrect value for 'vid' which indicates the mac filter's VLAN status. The current implementation for handling the removal of VLANs is wrong for a couple reasons. The first is that when i40e_vsi_kill_vlan iterates through the MAC filters, it fails to account for the MAC filter status; i.e. it's not accommodating for filters that are about to be deleted. The second problem is that MAC filters can be deleted in other places (specifically i40e_set_rx_mode). Thus if it occurs that all the VLAN MAC filters get deleted we need to switch out of VLAN mode, but the code path through i40e_vsi_kill_vlan has already been executed and we're now stuck in VLAN mode. This patch fixes the issue by removing the check from i40e_vsi_kill_vlan and puts the check instead in i40e_sync_vsi_filters where we're guaranteed to see all filter deletions and can properly detect when we need to switch out of VLAN mode. Change-ID: Ib38fe6034b356eee9a0e20b8a9eeed5ff2debcd9 Signed-off-by:
Alan Brady <alan.brady@intel.com> Tested-by:
-
Jacob Keller authored
Currently, we fail to correctly restore filters on the temporary add list when we fail to allocate memory either for deletion or addition. Replace calls to "goto out;" with calls to a new location that correctly handles memory allocation failures. Note that it is safe for us to call i40e_undo_filter_entries on the tmp_del_list even after we've deleted filters because at this point it will be empty, so we don't need to separate the logic for add and delete failure. Change-Id: Iee107fd219c6e03e2fd9645c2debf8e8384a8521 Signed-off-by:
-
Jacob Keller authored
Replace the mac_filter_list with a static size hash table of 8bits. The primary advantage of this is a decrease in latency of operations related to searching for specific MAC filters, including .set_rx_mode. Using a linked list resulted in several locations which were O(n^2). Using a hash table should give us latency growth closer to O(n*log(n)). Change-ID: I5330bd04053b880e670210933e35830b95948ebb Signed-off-by:
-
Jacob Keller authored
When inside a loop where we call i40e_del_filter we use an O(n^2) pattern where i40e_del_filter calls i40e_find_filter for us. We can avoid this O(n^2) logic by factoring a function, __i40e_del_filter() out from the i40e_del_filter code. This allows us to re-use the delete logic where appropriate without having to search for the filter twice. This new function benefits several functions including i40e_vsi_add_vlan, i40e_vsi_kill_vlan, i40e_del_mac_vlan_all, and i40e_vsi_release. Change-ID: I75fabe0f53bf73f56b80d342e5fdcfcc28f4d3eb Signed-off-by:
-
Jacob Keller authored
When adding new MAC address filters, the driver determines if it should behave in VLAN mode (where all MAC addresses get assigned to every existing VLAN) or in non-VLAN mode where MAC addresses get assigned the VLAN_ANY identifier. Under some circumstances it is possible that a VLAN has been marked for removal (such that all filters of that VLAN are set to I40E_FILTER_REMOVE), and a subsequent call to i40e_put_mac_in_vlan may occur prior to the driver subtask that syncs filters to the hardware. In this case, we may add filters to the new removed VLAN, even though it should have been removed. This is most obvious when first adding a new VLAN. We will delete all filters which are in I40E_VLAN_ANY (-1) and then re-add them as in VLAN 0 (untagged). Then before we sync filters, we will add new MAC address filter, which will be added to every VLAN that exists. Unfortunately, this will include I40E_VLAN_ANY, so we will end up incorrectly adding filters to the -1 VLAN. This can be fixed by simply skipping all filters which are marked for removal. A similar check is not necessary in i40e_del_mac_all_vlan, since we are deleting, and any filter which we find already marked for removal would simply be deleted again, which doesn't cause any issues. Change-Id: I7962154013ce02fe950584690aeeb3ed853d0086 Signed-off-by:
-
Jacob Keller authored
When a PVID has been assigned to a VSI, the function i40e_put_mac_in_vlan arbitrarily modifies all filters to have the same VLAN. This is obviously incorrect because it could be modifying active filters without putting them into the NEW state. The correct method is to remove then re-add filters which is already done in the code where we assign the PVID. Fix this issue and a few other minor nits at the same time. First, when we have a PVID don't even bother looping and simply add the filter with the PVID immediately. In the case of the loop, we now can remove several checks. We also don't need to use i40e_find_filter first before calling i40e_add_filter, since i40e_add_filter implicitly does a lookup already. Finally, update the return semantics of this function so that on failure to add a filter it returns NULL, but on success, it returns the last filter added. Otherwise, we're just returning the last filter in the list. An alternative fix might be to return 0 or an error code, but this is pretty invasive to every call site. Change-ID: I2325dfd843aec76d89fb0d7cb0e7c4f290a34840 Signed-off-by:
-
Jacob Keller authored
A future patch will be modifying these functions and making a call to a static function which currently is defined after these functions. Move them in a separate patch to ease review and ensure the moved code is correct. Change-ID: I2ca7fd4e10c0c07ed2291db1ea41bf5987fc6474 Signed-off-by:
-
Jacob Keller authored
The kernel provides __dev_uc_sync and __dev_mc_sync in order for drivers which need individual notification of add and delete for each filter. These functions allow us to vastly simplify our .set_rx_mode handler. We need to implement two functions for sync and unsync which add and remove filters respectively. This change avoids a very complex and inefficient algorithm which resulted in an abnormal latency for the .set_rx_mode NDO operation. The resulting code after this change is more readable, more efficient, and less code. Due to the callback signature used by these functions we also must update several other functions to take a const u8 * pointer. Change-Id: I2ca7fd4e10c0c07ed2291db1ea41bf5987fc6474 Signed-off-by:
-
