The hfi1 user SDMA pinned-page cache will leave a stale cache entry when
the cache-entry's virtual address range is invalidated but that cache
entry is in-use by an outstanding SDMA request.
Subsequent user SDMA requests with buffers in or spanning the virtual
address range of the stale cache entry will result in packets constructed
from the wrong memory, the physical pages pointed to by the stale cache
entry.
To fix this, remove mmu_rb_node cache entries from the mmu_rb_handler
cache independent of the cache entry's refcount. Add 'struct kref
refcount' to struct mmu_rb_node and manage mmu_rb_node lifetime with
kref_get() and kref_put().
mmu_rb_node.refcount makes sdma_mmu_node.refcount redundant. Remove
'atomic_t refcount' from struct sdma_mmu_node and change sdma_mmu_node
code to use mmu_rb_node.refcount.
Move the mmu_rb_handler destructor call after a
wait-for-SDMA-request-completion call so mmu_rb_nodes that need
mmu_rb_handler's workqueue to queue themselves up for destruction from an
interrupt context may do so.
Fixes: f48ad614c1 ("IB/hfi1: Move driver out of staging")
Fixes: 00cbce5cbf ("IB/hfi1: Fix bugs with non-PAGE_SIZE-end multi-iovec user SDMA requests")
Link: https://lore.kernel.org/r/168451393605.3700681.13493776139032178861.stgit@awfm-02.cornelisnetworks.com
Reviewed-by: Dean Luick <dean.luick@cornelisnetworks.com>
Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
hfi1 user SDMA request processing has two bugs that can cause data
corruption for user SDMA requests that have multiple payload iovecs
where an iovec other than the tail iovec does not run up to the page
boundary for the buffer pointed to by that iovec.a
Here are the specific bugs:
1. user_sdma_txadd() does not use struct user_sdma_iovec->iov.iov_len.
Rather, user_sdma_txadd() will add up to PAGE_SIZE bytes from iovec
to the packet, even if some of those bytes are past
iovec->iov.iov_len and are thus not intended to be in the packet.
2. user_sdma_txadd() and user_sdma_send_pkts() fail to advance to the
next iovec in user_sdma_request->iovs when the current iovec
is not PAGE_SIZE and does not contain enough data to complete the
packet. The transmitted packet will contain the wrong data from the
iovec pages.
This has not been an issue with SDMA packets from hfi1 Verbs or PSM2
because they only produce iovecs that end short of PAGE_SIZE as the tail
iovec of an SDMA request.
Fixing these bugs exposes other bugs with the SDMA pin cache
(struct mmu_rb_handler) that get in way of supporting user SDMA requests
with multiple payload iovecs whose buffers do not end at PAGE_SIZE. So
this commit fixes those issues as well.
Here are the mmu_rb_handler bugs that non-PAGE_SIZE-end multi-iovec
payload user SDMA requests can hit:
1. Overlapping memory ranges in mmu_rb_handler will result in duplicate
pinnings.
2. When extending an existing mmu_rb_handler entry (struct mmu_rb_node),
the mmu_rb code (1) removes the existing entry under a lock, (2)
releases that lock, pins the new pages, (3) then reacquires the lock
to insert the extended mmu_rb_node.
If someone else comes in and inserts an overlapping entry between (2)
and (3), insert in (3) will fail.
The failure path code in this case unpins _all_ pages in either the
original mmu_rb_node or the new mmu_rb_node that was inserted between
(2) and (3).
3. In hfi1_mmu_rb_remove_unless_exact(), mmu_rb_node->refcount is
incremented outside of mmu_rb_handler->lock. As a result, mmu_rb_node
could be evicted by another thread that gets mmu_rb_handler->lock and
checks mmu_rb_node->refcount before mmu_rb_node->refcount is
incremented.
4. Related to #2 above, SDMA request submission failure path does not
check mmu_rb_node->refcount before freeing mmu_rb_node object.
If there are other SDMA requests in progress whose iovecs have
pointers to the now-freed mmu_rb_node(s), those pointers to the
now-freed mmu_rb nodes will be dereferenced when those SDMA requests
complete.
Fixes: 7be85676f1 ("IB/hfi1: Don't remove RB entry when not needed.")
Fixes: 7724105686 ("IB/hfi1: add driver files")
Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Signed-off-by: Patrick Kelsey <pat.kelsey@cornelisnetworks.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Link: https://lore.kernel.org/r/168088636445.3027109.10054635277810177889.stgit@252.162.96.66.static.eigbox.net
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The following allocation with large txqueuelen will result in the
following warning:
Call Trace:
__alloc_pages_nodemask+0x283/0x2c0
kmalloc_large_node+0x3c/0xa0
__kmalloc_node+0x22a/0x2f0
hfi1_ipoib_txreq_init+0x19f/0x330 [hfi1]
hfi1_ipoib_setup_rn+0xd3/0x1a0 [hfi1]
rdma_init_netdev+0x5a/0x80 [ib_core]
ipoib_intf_init+0x6c/0x350 [ib_ipoib]
ipoib_intf_alloc+0x5c/0xc0 [ib_ipoib]
ipoib_add_one+0xbe/0x300 [ib_ipoib]
add_client_context+0x12c/0x1a0 [ib_core]
ib_register_client+0x147/0x190 [ib_core]
ipoib_init_module+0xdd/0x132 [ib_ipoib]
do_one_initcall+0x46/0x1c3
do_init_module+0x5a/0x220
load_module+0x14c5/0x17f0
__do_sys_init_module+0x13b/0x180
do_syscall_64+0x5b/0x1a0
entry_SYSCALL_64_after_hwframe+0x65/0xca
For ipoib, the txqueuelen is modified with the module parameter
send_queue_size.
Fix by changing to use kv versions of the same allocator to handle the
large allocations. The allocation embeds a hdr struct that is dma mapped.
Change that struct to a pointer to a kzalloced struct.
Cc: stable@vger.kernel.org
Fixes: d99dc602e2 ("IB/hfi1: Add functions to transmit datagram ipoib packets")
Link: https://lore.kernel.org/r/1642287756-182313-3-git-send-email-mike.marciniszyn@cornelisnetworks.com
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Merged due to dependencies in following patches.
Conflict in drivers/infiniband/hw/hfi1/ipoib_tx.c resolved by hand to take
the %p change and txq stats rename together.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This patch removes kmem cache allocation and deallocation in favor of
having the ipoib_txreq in the ring.
The consumer is now the packet sending side allocating tx descriptors from
ring and the producer is the napi interrupt handling freeing tx
descriptors.
The locks are now eliminated because the napi tx lock insures a single
consumer and the napi handling insures a single producer.
The napi poll is converted to memory poll looking for items that have been
marked completed.
Fixes: d99dc602e2 ("IB/hfi1: Add functions to transmit datagram ipoib packets")
Link: https://lore.kernel.org/r/20210913132826.131370.4397.stgit@awfm-01.cornelisnetworks.com
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Use netdev->tstats instead of a member of hfi1_ipoib_dev_priv for storing
a pointer to the per-cpu counters. This allows us to use core
functionality for statistics handling.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When running iperf in a two host configuration the following trace can
occur:
[ 319.728730] NETDEV WATCHDOG: ib0 (hfi1): transmit queue 0 timed out
The issue happens because the current implementation relies on the netif
txq being stopped to control the flushing of the tx list.
There are two resources that the transmit logic can wait on and stop the
txq:
- SDMA descriptors
- Ring space to hold completions
The ring space is tested on the sending side and relieved when the ring is
consumed in the napi tx reaping.
Unfortunately, that reaping can run conncurrently with the workqueue
flushing of the txlist. If the txq is started just before the workitem
executes, the txlist will never be flushed, leading to the txq being
stuck.
Fix by:
- Adding sleep/wakeup wrappers
* Use an atomic to control the call to the netif routines inside the
wrappers
- Use another atomic to record ring space exhaustion
* Only wakeup when the a ring space exhaustion has happened and it
relieved
Add additional wrappers to clarify the ring space resource handling.
Fixes: d99dc602e2 ("IB/hfi1: Add functions to transmit datagram ipoib packets")
Link: https://lore.kernel.org/r/20200623204327.108092.4024.stgit@awfm-01.aw.intel.com
Reviewed-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The current code mishandles -EBUSY in two ways:
- The flow change doesn't test the return from the flush and runs on to
process the current packet racing with the wakeup processing
- The -EBUSY handling for a single packet inserts the tx into the txlist
after the submit call, racing with the same wakeup processing
Fix the first by dropping the skb and returning NETDEV_TX_OK.
Fix the second by insuring the the list entry within the txreq is inited
when allocated. This enables the sleep routine to detect that the txreq
has used the non-list api and queue the packet to the txlist.
Both flaws can lead to having the flushing thread executing in causing two
threads to manipulate the txlist.
Fixes: d99dc602e2 ("IB/hfi1: Add functions to transmit datagram ipoib packets")
Link: https://lore.kernel.org/r/20200623204321.108092.83898.stgit@awfm-01.aw.intel.com
Reviewed-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This patch implements the mechanism to accelerate the transmit side of
a multiple transmit queue RDMA netdev by submitting the packets to
the SDMA engine directly instead of sending through the verbs layer.
This patch also changes the UD/SEND_ONLY op to output the entropy value
in byte 0 of deth[1]. UD/SEND_ONLY_WITH_IMMEDIATE uses the previous
behavior with no entropy value being output.
The code in the ipoib rdma netdev which submits tx requests upon
successful submission will call trace_sdma_output_ibhdr to output
the ibhdr to the trace buffer.
Link: https://lore.kernel.org/r/20200511160548.173205.45616.stgit@awfm-01.aw.intel.com
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Gary Leshner <Gary.S.Leshner@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
