nvmet_rdma_release_rsp() may free the response before using it at error
flow.

Fixes: 8407879c ("nvmet-rdma: fix possible bogus dereference under heavy load")
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This reverts commit 2acf70ad.

The commit never really fixed the intended issue and caused all
kinds of other issues, including a use before initialization.
Suggested-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Create a configfs attribute in each nvme-fabrics namespace to enable P2P
memory use.  The attribute may be enabled (with a boolean) or a specific
P2P device may be given (with the device's PCI name).

When enabled, the namespace will ensure the underlying block device
supports P2P and is compatible with any specified P2P device.  If no device
was specified it will ensure there is compatible P2P memory somewhere in
the system.  Enabling a namespace with P2P memory will fail with EINVAL
(and an appropriate dmesg error) if any of these conditions are not met.

Once a controller is set up on a specific port, the P2P device to use for
each namespace will be found and stored in a radix tree by namespace ID.
When memory is allocated for a request, the tree is used to look up the P2P
device to allocate memory against.  If no device is in the tree (because no
appropriate device was found), or if allocation of P2P memory fails, fall
back to using regular memory.
Signed-off-by: NStephen Bates <sbates@raithlin.com>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
[hch: partial rewrite of the initial code]
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NLogan Gunthorpe <logang@deltatee.com>
Signed-off-by: NBjorn Helgaas <bhelgaas@google.com>

Add helpers to allocate and free the SGL in a struct nvmet_req:

  int nvmet_req_alloc_sgl(struct nvmet_req *req)
  void nvmet_req_free_sgl(struct nvmet_req *req)

This will be expanded in a future patch to implement peer-to-peer memory
DMAs and should be common with all target drivers.

The new helpers are used in nvmet-rdma.  Seeing we use req.transfer_len as
the length of the SGL it is set earlier and cleared on any error.  It also
seems to be unnecessary to accumulate the length as the map_sgl functions
should only ever be called once per request.
Signed-off-by: NLogan Gunthorpe <logang@deltatee.com>
Signed-off-by: NBjorn Helgaas <bhelgaas@google.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NSagi Grimberg <sagi@grimberg.me>

This patch avoids that sparse complains about missing declarations.
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Reviewed-by: NChaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Queue deletion is done asynchronous when the last reference on the queue
is dropped.  Thus, in order to make sure we don't over allocate under a
connect/disconnect storm, we let queue deletion complete before making
forward progress.

However, given that we flush the system_wq from rdma_cm context which
runs from a workqueue context, we can have a circular locking complaint
[1]. Fix that by using a private workqueue for queue deletion.

[1]:
======================================================
WARNING: possible circular locking dependency detected
4.19.0-rc4-dbg+ #3 Not tainted
------------------------------------------------------
kworker/5:0/39 is trying to acquire lock:
00000000a10b6db9 (&id_priv->handler_mutex){+.+.}, at: rdma_destroy_id+0x6f/0x440 [rdma_cm]

but task is already holding lock:
00000000331b4e2c ((work_completion)(&queue->release_work)){+.+.}, at: process_one_work+0x3ed/0xa20

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #3 ((work_completion)(&queue->release_work)){+.+.}:
       process_one_work+0x474/0xa20
       worker_thread+0x63/0x5a0
       kthread+0x1cf/0x1f0
       ret_from_fork+0x24/0x30

-> #2 ((wq_completion)"events"){+.+.}:
       flush_workqueue+0xf3/0x970
       nvmet_rdma_cm_handler+0x133d/0x1734 [nvmet_rdma]
       cma_ib_req_handler+0x72f/0xf90 [rdma_cm]
       cm_process_work+0x2e/0x110 [ib_cm]
       cm_req_handler+0x135b/0x1c30 [ib_cm]
       cm_work_handler+0x2b7/0x38cd [ib_cm]
       process_one_work+0x4ae/0xa20
nvmet_rdma:nvmet_rdma_cm_handler: nvmet_rdma: disconnected (10): status 0 id 0000000040357082
       worker_thread+0x63/0x5a0
       kthread+0x1cf/0x1f0
       ret_from_fork+0x24/0x30
nvme nvme0: Reconnecting in 10 seconds...

-> #1 (&id_priv->handler_mutex/1){+.+.}:
       __mutex_lock+0xfe/0xbe0
       mutex_lock_nested+0x1b/0x20
       cma_ib_req_handler+0x6aa/0xf90 [rdma_cm]
       cm_process_work+0x2e/0x110 [ib_cm]
       cm_req_handler+0x135b/0x1c30 [ib_cm]
       cm_work_handler+0x2b7/0x38cd [ib_cm]
       process_one_work+0x4ae/0xa20
       worker_thread+0x63/0x5a0
       kthread+0x1cf/0x1f0
       ret_from_fork+0x24/0x30

-> #0 (&id_priv->handler_mutex){+.+.}:
       lock_acquire+0xc5/0x200
       __mutex_lock+0xfe/0xbe0
       mutex_lock_nested+0x1b/0x20
       rdma_destroy_id+0x6f/0x440 [rdma_cm]
       nvmet_rdma_release_queue_work+0x8e/0x1b0 [nvmet_rdma]
       process_one_work+0x4ae/0xa20
       worker_thread+0x63/0x5a0
       kthread+0x1cf/0x1f0
       ret_from_fork+0x24/0x30

Fixes: 777dc823 ("nvmet-rdma: occasionally flush ongoing controller teardown")
Reported-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Tested-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Currently we always repost the recv buffer before we send a response
capsule back to the host. Since ordering is not guaranteed for send
and recv completions, it is posible that we will receive a new request
from the host before we got a send completion for the response capsule.

Today, we pre-allocate 2x rsps the length of the queue, but in reality,
under heavy load there is nothing that is really preventing the gap to
expand until we exhaust all our rsps.

To fix this, if we don't have any pre-allocated rsps left, we dynamically
allocate a rsp and make sure to free it when we are done. If under memory
pressure we fail to allocate a rsp, we silently drop the command and
wait for the host to retry.
Reported-by: NSteve Wise <swise@opengridcomputing.com>
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
[hch: dropped a superflous assignment]
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Instead of declaring and passing a dummy 'bad_wr' pointer, pass NULL
as third argument to ib_post_(send|recv|srq_recv)().
Signed-off-by: NBart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Posting receive buffer operation can fail, thus we should make
sure to have an error flow during initialization phase. While
we're here, add a debug print in case of a failure.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

ib_post_send operation should succeed unless something unusual
happened to the ib device.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

The patch enables inline data sizes using up to 4 recv sges, and capping
the size at 16KB or at least 1 page size.  So on a 4K page system, up to
16KB is supported, and for a 64K page system 1 page of 64KB is supported.

We avoid > 0 order page allocations for the inline buffers by using
multiple recv sges, one for each page.  If the device cannot support
the configured inline data size due to lack of enough recv sges, then
log a warning and reduce the inline size.

Add a new configfs port attribute, called param_inline_data_size,
to allow configuring the size of inline data for a given nvmf port.
The maximum size allowed is still enforced by nvmet-rdma with
NVMET_RDMA_MAX_INLINE_DATA_SIZE, which is now max(16KB, PAGE_SIZE).
And the default size, if not specified via configfs, is still PAGE_SIZE.
This preserves the existing behavior, but allows larger inline sizes
for small page systems.  If the configured inline data size exceeds
NVMET_RDMA_MAX_INLINE_DATA_SIZE, a warning is logged and the size is
reduced.  If param_inline_data_size is set to 0, then inline data is
disabled for that nvmf port.
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

This patch replaces the ib_device_attr.max_sge with max_send_sge and
max_recv_sge. It allows ulps to take advantage of devices that have very
different send and recv sge depths.  For example cxgb4 has a max_recv_sge
of 4, yet a max_send_sge of 16.  Splitting out these attributes allows
much more efficient use of the SQ for cxgb4 with ulps that use the RDMA_RW
API. Consider a large RDMA WRITE that has 16 scattergather entries.
With max_sge of 4, the ulp would send 4 WRITE WRs, but with max_sge of
16, it can be done with 1 WRITE WR.
Acked-by: NSagi Grimberg <sagi@grimberg.me>
Acked-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Acked-by: NShiraz Saleem <shiraz.saleem@intel.com>
Acked-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

The .remove_one function is called for any ib_device removal.
In case the removed device has no reference in our driver, there
is no need to flush the system work queue.
Reviewed-by: NIsrael Rukshin <israelr@mellanox.com>
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We free nvmet rdma queues while handling rdma_cm events.
In order to avoid this we destroy the qp and the queue after destroying
the cm_id which guarantees that all rdma_cm events are done.
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Its perfectly valid to assign a nvmet port to listen on "any"
IP address (traddr 0.0.0.0 for ipv4 address family) for IP based
transport ports. However, we must not return this address in
discovery log entries. Instead we need to return the address
where the request was accepted on (req->port address).

Since this is nvme transport specific, introduce an optional
.disc_traddr interface that is designed to check that a
port in question is bound to "any" IP address and if so, set
the traddr from the port where the request came from.
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

It is a bit chatty to report on every deleted queue, so keep it for debug
purposes only.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

We already do that when we are notified in device removal
which is triggered when unregistering as an ib client.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Use the sgl_alloc() and sgl_free() functions instead of open coding
these functions.
Signed-off-by: NBart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Currently the NVMe target stores the expexted data length in req->data_len
and uses that for data transfer decisions, but that does not take the
actual transfer length in the SGLs into account.  So this adds a new
transfer_len field, into which the transport drivers store the actual
transfer length.  We then check the two match before actually executing
the command.

The FC transport driver already had such a field, which is removed in
favour of the common one.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

A NULL deref happens when nvmet_rdma_remove_one() is called more than once
(e.g. while connected via 2 ports).
The first call frees the queues related to the first ib_device but
doesn't remove them from the queue list.
While calling nvmet_rdma_remove_one() for the second ib_device it goes over
the full queue list again and we get the NULL deref.

Fixes: f1d4ef7d ("nvmet-rdma: register ib_client to not deadlock in device removal")
Signed-off-by: NIsrael Rukshin <israelr@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NSagi Grimberg <sagi@grmberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Now that its not needed, we can simply not assign it.
Reviewed-by: NJohannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

We can deadlock in case we got to a device removal
event on a queue which is already in the process of
destroying the cm_id is this is blocking until all
events on this cm_id will drain. On the other hand
we cannot guarantee that rdma_destroy_id was invoked
as we only have indication that the queue disconnect
flow has been queued (the queue state is updated before
the realease work has been queued).

So, we leave all the queue removal to a separate ib_client
to avoid this deadlock as ib_client device removal is in
a different context than the cm_id itself.
Reported-by: NShiraz Saleem <shiraz.saleem@intel.com>
Tested-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

No need to differentiate fabrics from pci/loop, also lower
it to 32 as we don't really need 256 inflight admin commands.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NMartin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NKeith Busch <keith.busch@intel.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

On rdma read errors, release the sq ref that was taken
when the req was initialized. This avoids a hang in
nvmet_sq_destroy() when the queue is being freed.
Signed-off-by: NVijay Immanuel <vijayi@attalasystems.com>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

Instead of parsing address strings, use a generic
helper. This also adds ipv6 (with address scopes)
support.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

If we are attacked with establishments/teradowns we need to
make sure we do not consume too much system memory. Thus
let ongoing controller teardowns complete before accepting
new controller establishments.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

When handling a new recv command, we grab a new rsp resource and
check for the queue state being live. In case the queue is not in
live state, we simply restore the rsp back to the free list. However
in this flow we didn't set rsp->queue yet, so we cannot dereference it.

Instead, make sure to initialize rsp->queue (and other rsp members)
as soon as possible so we won't reference uninitialized variables.
Reported-by: NYi Zhang <yizhan@redhat.com>
Reported-by: NRaju Rangoju <rajur@chelsio.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NRaju Rangoju <rajur@chelsio.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

When handling a new recv command, we grab a new rsp resource and
check for the queue state being live. In case the queue is not in
live state, we simply restore the rsp back to the free list. However
in this flow we didn't set rsp->queue yet, so we cannot dereference it.

Instead, make sure to initialize rsp->queue (and other rsp members)
as soon as possible so we won't reference uninitialized variables.
Reported-by: NYi Zhang <yizhan@redhat.com>
Reported-by: NRaju Rangoju <rajur@chelsio.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NRaju Rangoju <rajur@chelsio.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

According to the preceeding goto, it is likely that 'out_destroy_sq' was
expected here.
Signed-off-by: NChristophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

Also remove redundant debug prints.
Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Reviewed-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NJens Axboe <axboe@fb.com>

This patch performs dma sync operations on nvme_command
and nvme_completion.

nvme_command is synced
(a) on receiving of the recv queue completion for cpu access.
(b) before posting recv wqe back to rdma adapter for device access.

nvme_completion is synced
(a) on receiving of the recv queue completion of associated
nvme_command for cpu access.
(b) before posting send wqe to rdma adapter for device access.

This patch is generated for git://git.infradead.org/nvme-fabrics.git
Branch: nvmf-4.10
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMax Gurtovoy <maxg@mellanox.com>

Acked-by: NSagi Grimberg <sagi@grimberg.me>
Signed-off-by: NSteve Wise <swise@opengridcomputing.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Signed-off-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

nvmet_sq_init() returns a value <= 0. nvmet_rdma_cm_reject() expects
a second argument that is a NVME_RDMA_CM_* constant. Hence this patch.
Signed-off-by: NBart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: NSagi Grimberg <sagi@grimbeg.me>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

draining the qp right after disconnect might not suffice because
the nvmet sq is not fully drained (in nvmet_sq_destroy) and we might
see completions after the drain. Instead, drain right before the
qp destroy which comes after the sq destruction and we can be sure
that no posts come after the drain.
Tested-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

In case we accepted a queue connection and it failed, we might not
remove the queue from the list until we unload and clean it up.
We should delete it from the queue list on the relevant handler.
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

When we initiate queue teardown sequence we call rdma_destroy_qp
which clears cm_id->qp, afterwards we call rdma_destroy_id, but
we might see a rdma_cm event in between with a cleared cm_id->qp
so watch out for that and silently ignore the event because this
means that the queue teardown sequence is in progress.
Signed-off-by: NBart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>

Instead of exposing ib_get_dma_mr to ULPs and letting them use it more or
less unchecked, this moves the capability of creating a global rkey into
the RDMA core, where it can be easily audited.  It also prints a warning
everytime this feature is used as well.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NJason Gunthorpe <jgunthorpe@obsidianresearch.com>
Reviewed-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>