If ib_umem_release() is called after ib_uverbs_close() sets context->closing,
then a process can get stuck in a D state, because the code boils down to
if (down_write_trylock(&mm->mmap_sem))
down_write(&mm->mmap_sem);
which is obviously a stupid instant deadlock. Fix the code so that we
only try to take the lock once.
This bug was introduced in commit f7c6a7b5 ("IB/uverbs: Export
ib_umem_get()/ib_umem_release() to modules") which fortunately never
made it into a release, and was reported by Pete Wyckoff <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
Inline data segments in send WQEs are not allowed to cross a 64 byte
boundary. We use inline data segments to hold the UD headers for MLX
QPs (QP0 and QP1). A send with GRH on QP1 will have a UD header that
is too big to fit in a single inline data segment without crossing a
64 byte boundary, so split the header into two inline data segments.
Signed-off-by: Roland Dreier <[email protected]>
Upcoming firmware introduces command interface revision 3, which
changes the way port capabilities are queried and set. Update the
driver to handle both the new and old command interfaces by adding a
new MLX4_FLAG_OLD_PORT_CMDS that it is set after querying the firmware
interface revision and then using the correct interface based on the
setting of the flag.
Signed-off-by: Roland Dreier <[email protected]>
The calculation of max_inline_data in set_kernel_sq_size() is bogus,
since it doesn't take into account the fact that inline segments may
not cross a 64-byte boundary, and hence multiple inline segments will
probably need to be used to post large inline sends.
We don't support inline sends for kernel QPs anyway, so there's no
point in doing this calculation anyway, since the field is just zeroed
out a little later. So just delete the bogus calculation.
Signed-off-by: Roland Dreier <[email protected]>
New ConnectX firmware introduces FW command interface revision 2,
which requires that for each QP, a chunk of send queue entries (the
"headroom") is kept marked as invalid, so that the HCA doesn't get
confused if it prefetches entries that haven't been posted yet. Add
code to the driver to do this, and also update the user ABI so that
userspace can request that the prefetcher be turned off for userspace
QPs (we just leave the prefetcher on for all kernel QPs).
Unfortunately, marking send queue entries this way is confuses older
firmware, so we change the driver to allow only FW command interface
revisions 2. This means that users will have to update their firmware
to work with the new driver, but the firmware is changing quickly and
the old firmware has lots of other bugs anyway, so this shouldn't be too
big a deal.
Based on a patch from Jack Morgenstein <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
Doing max(1, foo) where foo is u32 generates a warning, because 1 is a
signed constant. Fix this by using 1U instead.
Signed-off-by: Roland Dreier <[email protected]>
Cast the increment added to wq->tail when send completions are
processed to u16 to avoid using wrong values caused by standard
integer promotions.
The same bug was fixed in libmlx4 by Eli Cohen <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
QPs attached to an SRQ must never have their own RQ, and QPs not
attached to SRQs must have an RQ with at least 1 entry. Enforce all
of this in set_rq_size().
Based on a patch by Eli Cohen <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
HCA firmware with incompatible changes to the FW commmand interface is
coming soon. Add a check of the interface revision during
initialization and bail out if the firmware advertises a revision that
the driver doesn't know about. This will avoid strange failures later
if the driver goes on using the wrong interface revision.
Signed-off-by: Roland Dreier <[email protected]>
We need to pass the same dev_id to free_irq() and request_irq(). When
using MSI-X, the MLX4_EQ_CATAS interrupt uses a different dev_id from
the other interrupts.
Signed-off-by: Roland Dreier <[email protected]>
We may call mlx4_dispatch_event() before mlx4_register_device() is
called for a device, because for example a catastrophic error happens
immediately after we enable interrupts. Therefore priv->ctx_list and
priv->ctx_lock need to be initialized earlier.
This bug was actually exposed by the MSI-X bug that returned IRQ numbers
to drivers in reverse order, so that the first FW command
interrupt looked to mlx4 like a catastrophic error.
Signed-off-by: Roland Dreier <[email protected]>
Set last allocated object to the object after the one just allocated
before ORing in the extra top bits. Also handle the case where this
wraps around.
Signed-off-by: Roland Dreier <[email protected]>
If a QP is attached to a shared receive queue (SRQ), then it doesn't
have a receive queue (RQ). So don't allocate an RQ doorbell (or map a
doorbell from userspace for userspace QPs) for that QP.
Signed-off-by: Roland Dreier <[email protected]>
wr->opcode is invalid if it's >= ARRAY_SIZE(mlx4_ib_opcode), not just
strictly >.
This was spotted by the Coverity checker (CID 1643).
Signed-off-by: Roland Dreier <[email protected]>
Don't overrun fname[] array when decoding device flags.
This was spotted by the Coverity checker (CID 1642).
Signed-off-by: Roland Dreier <[email protected]>
Clean up ib_query_port() and ib_modify_port() slightly by using the
just-added start_port() and end_port() helpers.
Signed-off-by: Roland Dreier <[email protected]>
Lockdep found the following potential deadlock between mcast_lock and
n_mcast_grps_lock: mcast_lock is taken from both interrupt context and
process context, so spin_lock_irqsave() must be used to take it.
n_mcast_grps_lock is only taken from process context, so at first it
seems safe to take it with plain spin_lock(); however, it also nests
inside mcast_lock, and hence we could deadlock:
cpu A cpu B
ipath_mcast_add():
spin_lock_irq(&mcast_lock);
ipath_mcast_detach():
spin_lock(&n_mcast_grps_lock);
<enter interrupt>
ipath_mcast_find():
spin_lock_irqsave(&mcast_lock);
spin_lock(&n_mcast_grps_lock);
Fix this by using spin_lock_irq() to take n_mcast_grps_lock.
Signed-off-by: Roland Dreier <[email protected]>
Add an InfiniBand driver for Mellanox ConnectX adapters. Because
these adapters can also be used as ethernet NICs and Fibre Channel
HBAs, the driver is split into two modules:
mlx4_core: Handles low-level things like device initialization and
processing firmware commands. Also controls resource allocation
so that the InfiniBand, ethernet and FC functions can share a
device without stepping on each other.
mlx4_ib: Handles InfiniBand-specific things; plugs into the
InfiniBand midlayer.
Signed-off-by: Roland Dreier <[email protected]>
When memory pinned with ib_umem_get() is released, ib_umem_release()
needs to subtract the amount of memory being unpinned from
mm->locked_vm. However, ib_umem_release() may be called with
mm->mmap_sem already held for writing if the memory is being released
as part of an munmap() call, so it is sometimes necessary to defer
this accounting into a workqueue.
However, the work struct used to defer this accounting is dynamically
allocated before it is queued, so there is the possibility of failing
that allocation. If the allocation fails, then ib_umem_release has no
choice except to bail out and leave the process with a permanently
elevated locked_vm.
Fix this by allocating the structure to defer accounting as part of
the original struct ib_umem, so there's no possibility of failing a
later allocation if creating the struct ib_umem and pinning memory
succeeds.
Signed-off-by: Roland Dreier <[email protected]>
Export ib_umem_get()/ib_umem_release() and put low-level drivers in
control of when to call ib_umem_get() to pin and DMA map userspace,
rather than always calling it in ib_uverbs_reg_mr() before calling the
low-level driver's reg_user_mr method.
Also move these functions to be in the ib_core module instead of
ib_uverbs, so that driver modules using them do not depend on
ib_uverbs.
This has a number of advantages:
- It is better design from the standpoint of making generic code a
library that can be used or overridden by device-specific code as
the details of specific devices dictate.
- Drivers that do not need to pin userspace memory regions do not
need to take the performance hit of calling ib_mem_get(). For
example, although I have not tried to implement it in this patch,
the ipath driver should be able to avoid pinning memory and just
use copy_{to,from}_user() to access userspace memory regions.
- Buffers that need special mapping treatment can be identified by
the low-level driver. For example, it may be possible to solve
some Altix-specific memory ordering issues with mthca CQs in
userspace by mapping CQ buffers with extra flags.
- Drivers that need to pin and DMA map userspace memory for things
other than memory regions can use ib_umem_get() directly, instead
of hacks using extra parameters to their reg_phys_mr method. For
example, the mlx4 driver that is pending being merged needs to pin
and DMA map QP and CQ buffers, but it does not need to create a
memory key for these buffers. So the cleanest solution is for mlx4
to call ib_umem_get() in the create_qp and create_cq methods.
Signed-off-by: Roland Dreier <[email protected]>
Convert the IP-over-InfiniBand network device driver over to using
NAPI to handle completions for the main CQ. This covers all receives
as well as datagram mode sends; send completions for connected mode
connections are still handled from interrupt context.
Signed-off-by: Roland Dreier <[email protected]>
The semantics defined by the InfiniBand specification say that
completion events are only generated when a completions is added to a
completion queue (CQ) after completion notification is requested. In
other words, this means that the following race is possible:
while (CQ is not empty)
ib_poll_cq(CQ);
// new completion is added after while loop is exited
ib_req_notify_cq(CQ);
// no event is generated for the existing completion
To close this race, the IB spec recommends doing another poll of the
CQ after requesting notification.
However, it is not always possible to arrange code this way (for
example, we have found that NAPI for IPoIB cannot poll after
requesting notification). Also, some hardware (eg Mellanox HCAs)
actually will generate an event for completions added before the call
to ib_req_notify_cq() -- which is allowed by the spec, since there's
no way for any upper-layer consumer to know exactly when a completion
was really added -- so the extra poll of the CQ is just a waste.
Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for
ib_req_notify_cq() so that it can return a hint about whether the a
completion may have been added before the request for notification.
The return value of ib_req_notify_cq() is extended so:
< 0 means an error occurred while requesting notification
== 0 means notification was requested successfully, and if
IB_CQ_REPORT_MISSED_EVENTS was passed in, then no
events were missed and it is safe to wait for another
event.
> 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was
passed in. It means that the consumer must poll the
CQ again to make sure it is empty to avoid the race
described above.
We add a flag to enable this behavior rather than turning it on
unconditionally, because checking for missed events may incur
significant overhead for some low-level drivers, and consumers that
don't care about the results of this test shouldn't be forced to pay
for the test.
Signed-off-by: Roland Dreier <[email protected]>
There are quite a few places in ipoib_cm.c where we know IRQs are
enabled because we do something that sleeps in the same function, so
we can convert several occurrences of spin_lock_irqsave() to a plain
spin_lock_irq(). This cleans up the source a little and makes the
code smaller too:
add/remove: 0/0 grow/shrink: 1/5 up/down: 3/-51 (-48)
function old new delta
ipoib_cm_tx_reap 403 406 +3
ipoib_cm_stale_task 146 145 -1
ipoib_cm_dev_stop 173 172 -1
ipoib_cm_tx_handler 964 956 -8
ipoib_cm_rx_handler 956 937 -19
ipoib_cm_skb_reap 212 190 -22
Signed-off-by: Roland Dreier <[email protected]>
mthca_free_qp() already has local variables to hold the QP's send_cq
and recv_cq, so we can slightly clean up the calls to mthca_cq_clean()
by using those local variables instead of expressions like
to_mcq(qp->ibqp.send_cq).
Also, by cleaning the recv_cq first, we can avoid worrying about
whether the QP is attached to an SRQ for the second call, because we
would only clean send_cq if send_cq is not equal to recv_cq, and that
means send_cq cannot have any receive completions from the QP being
destroyed.
All this work even improves the generated code a bit:
add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-5 (-5)
function old new delta
mthca_free_qp 510 505 -5
Signed-off-by: Roland Dreier <[email protected]>
Commit b2875d4c ("IB/mthca: Always fill MTTs from CPU") causes a crash
in mthca_write_mtt() with non-memfree HCAs that have their memory
hidden (that is, have only two PCI BARs instead of having a third BAR
that allows access to the RAM attached to the HCA) on 64-bit
architectures. This is because the commit just before, c20e20ab
("IB/mthca: Merge MR and FMR space on 64-bit systems") makes
dev->mr_table.fmr_mtt_buddy equal to &dev->mr_table.mtt_buddy and
hence mthca_write_mtt() tries to write directly into the HCA's MTT
table. However, since that table is in the HCA's memory, this is
impossible without the PCI BAR that gives access to that memory.
This causes a crash because mthca_tavor_write_mtt_seg() basically
tries to dereference some offset of a NULL pointer. Fix this by
adding a test of MTHCA_FLAG_FMR in mthca_write_mtt() so that we always
use the WRITE_MTT firmware command rather than writing directly if
FMRs are not enabled.
Signed-off-by: Roland Dreier <[email protected]>
There's no point in printing the opcode field in the completion
handling debugging output, since the type of completion is already
printed at the beginning of the line. In fact the opcode field is not
even defined for completions with a status other than success.
Signed-off-by: Roland Dreier <[email protected]>
Since commit b1c1b6a3 ("IB/ipath: merge ipath_core and ib_ipath
drivers"), CONFIG_IPATH_CORE no longer exists, so there's no reason to
have a line for it in drivers/Makefile.
Pointed out by Robert P. J. Day <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
The garbled logic in mthca_alloc_memfree() causes it to return 0, even
if it fails to allocate all doorbell records. Fix it to return -ENOMEM
when it fails.
Signed-off-by: Roland Dreier <[email protected]>
An asynchronous event carries the port number that the event occurred
on, so there's no reason for an IPoIB interface to process an event
associated with a different local HCA port.
Signed-off-by: Roland Dreier <[email protected]>
If path_rec_completion() is passed a non-NULL path record pointer
along with an unsuccessful status value, the tracing code incorrectly
prints the (invalid) DLID from the path record rather than the more
interesting status code. The actual logic of the function correctly
uses the path record only if the status indicates a successful lookup.
Signed-off-by: Roland Dreier <[email protected]>
If no matching PD is found in ib_uverbs_reg_mr(), then the function
jumps to err_release without setting the return value ret. This means
that ret will hold the return value of the call to ib_umem_get() a few
lines earlier; if the function reaches the point where it looks for
the PD, we know that ib_umem_get() must have returned 0, so
ib_uverbs_reg_mr() ends up return 0 for a bad PD ID. Fix this by
setting ret to -EINVAL before jumping to the exit path when no PD is
found.
Signed-off-by: Roland Dreier <[email protected]>
Now that low-level drivers handle the conversion from an absolute rate
to a relative rate, there's no need for the IPoIB driver to keep track
of the local port's data rate.
Signed-off-by: Roland Dreier <[email protected]>
The static rate from the path record should be put into the address
vector -- a long time ago the rate in the address attributes needed to
be a relative rate, which required more munging, but now that the
conversion from absolute to relative is done in the low-level driver,
it's easy for ib_init_ah_from_path() to put the absolute rate in.
Cc: Jason Gunthorpe <[email protected]>
Cc: Sean Hefty <[email protected]>
Signed-off-by: Roland Dreier <[email protected]>
The change to allow allocating ICM chunks from coherent memory did not
increment the count of sg entries properly, so a chunk that required
more than allocation would not be mapped properly by the HCA.
Fix this by adding the missing increment of chunk->nsg.
Signed-off-by: Roland Dreier <[email protected]>
When clearing the ib_ah_attr parameter in to_ib_ah_attr(), use sizeof
*ib_ah_attr instead of sizeof *path.
Pointed out by Jack Morgenstein <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
mthca_device_mutex() can be initialized automatically with
DEFINE_MUTEX() rather than explicitly calling mutex_init(). This
saves a bit of text and shrinks the source by a line, so we may as
well do it....
Signed-off-by: Roland Dreier <[email protected]>
struct srp_device.fmr_page_mask was unsigned long, which means that
the top part of addresses above 4G was being chopped off on 32-bit
architectures. Of course nothing good happens when data from SRP
targets is DMAed to the wrong place.
Fix this by changing fmr_page_mask to u64, to match the addresses
actually used by IB devices.
Thanks to Brian Cain <[email protected]> and David McMillen
<[email protected]> for help diagnosing the bug and testing
the fix.
Signed-off-by: Roland Dreier <[email protected]>
The ib_dma_alloc_coherent() wrapper uses a u64* for the dma_handle
parameter, unlike dma_alloc_coherent, which uses dma_addr_t*. This
means that we need a temporary variable to handle the case when
ib_dma_alloc_coherent() just falls through directly to
dma_alloc_coherent() on architectures where sizeof u64 != sizeof
dma_addr_t.
Signed-off-by: Roland Dreier <[email protected]>
Move the initialization of ipoib_neigh's skb_queue into
ipoib_neigh_alloc(), since commit 2745b5b7 ("IPoIB: Fix skb leak when
freeing neighbour") will make iterate over the skb_queue to free any
packets left over when freeing the ipoib_neigh structure.
This fixes a crash when freeing ipoib_neigh structures allocated in
ipoib_mcast_send(), which otherwise don't have their skb_queue
initialized.
Signed-off-by: Roland Dreier <[email protected]>
Commit 51f65ebc ("IB/ipath - program intconfig register using new HT
irq hook"), which fixed interrupts for HyperTransport HCAs, broke PCI
Express HCAs, because for those HCAs, the driver uses the value of
pdev->irq before pci_enable_msi() and ends up getting a totally bogus
IRQ number. Fix this by using the value of pdev->irq after
pci_enable_msi().
Signed-off-by: Roland Dreier <[email protected]>
Remove variables that are set but then never looked at in the iSER
initiator. These cleanups came from David Binderman's list of "set
but never used" warnings from icc.
Signed-off-by: Roland Dreier <[email protected]>
Remove variables that are set but then never looked at in the ipath
driver. These cleanups came from David Binderman's list of "set but
never used" warnings from icc.
Signed-off-by: Roland Dreier <[email protected]>
ib_flush_fmr_pool() stashes away the request generation number
properly, but then goes ahead and rereads it every time it tests
whether the flush generation number has caught up. This means that
there is a theoretical possibility of livelock, if the request
generation number keeps getting bumped and the flush generation number
never catches up. The fix is simple: use the request generation
number read at the beginning of the function.
Also, atomic_inc() followed by atomic_read() can be replaced with
atomic_int_return(). There's no real requirement for atomicity here
but we might as well shrink the code.
This bug was discovered using David Binderman's list of "set but never
used" warnings from icc.
Signed-off-by: Roland Dreier <[email protected]>
The amso1100 driver was missing a couple of __devinit/__devexit
annotations for init/cleanup functions that are called from
__devinit/__devexit functions.
Reported by Randy Dunlap <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
Commit b3b30f5e ("IB/mthca: Recover from catastrophic errors")
introduced some section mismatch breakage, because the error recovery
code tears down and reinitializes the device, which calls into lots of
code originally marked __devinit and __devexit from regular .text.
Fix this by getting rid of these now-incorrect section markers.
Reported by Randy Dunlap <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
When ib_cancel_mad() is called, it puts the canceled send on a list
and schedules a "flushed" callback from process context. However,
this leaves a window where a receive completion could be processed
before the send is fully flushed.
This is fine, except that ib_find_send_mad() will find the MAD and
return it to the receive processing, which results in the sender
getting both a successful receive and a "flushed" send completion for
the same request. Understandably, this confuses the sender, which is
expecting only one of these two callbacks, and leads to grief such as
a use-after-free in IPoIB.
Fix this by changing ib_find_send_mad() to return a send struct only
if the status is still successful (and not "flushed"). The search of
the send_list already had this check, so this patch just adds the same
check to the search of the wait_list.
Signed-off-by: Roland Dreier <[email protected]>
Since pr_debug() has changed from a macro to an inline function when
DEBUG is not defined, its arguments now need to be defined even when
debugging is off. Therefore to_event_str() and to_qp_state_str() need
to be moved out of #ifdef DEBUG. The compiler will throw the
definitions away if DEBUG is not defined, but it needs to be able to
see that the functions exist.
Signed-off-by: Roland Dreier <[email protected]>
If the allocation of mr fails, then c2_reg_phys_mr() leaks the
page_list array it allocated earlier.
This was Coverity CID #1413.
Signed-off-by: Roland Dreier <[email protected]>
Make sure all 64-bit quantities are cast to unsigned long long
when printed with "%ll" printk formats.
Signed-off-by: Roland Dreier <[email protected]>
RFC 4391 ("Transmission of IP over InfiniBand (IPoIB)") says:
If the IB multicast group does not already exist, one must be
created first with the IPoIB link MTU. The MGID MUST use the same
P_Key, Q_Key, SL, MTU, and HopLimit as those used in the
broadcast-GID. The rest of attributes SHOULD follow the values used
in the broadcast-GID as well.
However, the current IPoIB driver is only setting the attributes
required by the InfiniBand spec to create a multicast group, so in
particular the MTU and HopLimit are not being set. Add these
attributes when creating MCGs, and also set the Rate attribute, since
IPoIB pays attention to that attribute as well.
Signed-off-by: Roland Dreier <[email protected]>
If a QP has separate send and receive CQs, then the send CQ will never
have receive completions from that QP in it. So when cleaning the
send CQ, there's no need to pass in an SRQ pointer, even if the QP is
attached to an SRQ.
Signed-off-by: Roland Dreier <[email protected]>
Split up ipoib_ib_handle_wc() into ipoib_ib_handle_rx_wc() and
ipoib_ib_handle_tx_wc() to make the code easier to read. This will
also help implement NAPI in the future.
Signed-off-by: Roland Dreier <[email protected]>
Remove some trailing whitespace that has snuck in despite the best
efforts of whitespace=error-all. Also fix a few other whitespace
bogosities.
Signed-off-by: Roland Dreier <[email protected]>
Lockdep warns when userspace creates a QP that uses different CQs for
send completions and receive completions, because both CQs are locked
and their mutexes belong to the same lock class. However, we know
that the mutexes are distinct and the nesting is safe (there is no
possibility of AB-BA deadlock because the mutexes are locked with
down_read()), so annotate the situation with SINGLE_DEPTH_NESTING to
get rid of the lockdep warning.
Signed-off-by: Roland Dreier <[email protected]>
There were two functions that open-coded idr_read_cq() in terms of
idr_read_uobj() rather than using the helper.
Signed-off-by: Roland Dreier <[email protected]>
It is supposed to be OK to call mthca_create_ah() and mthca_destroy_ah()
from any context. However, for mem-full HCAs, these functions use the
mthca_alloc() and mthca_free() bitmap helpers, and those helpers use
non-IRQ-safe spin_lock() internally. Lockdep correctly warns that
this could lead to a deadlock. Fix this by changing mthca_alloc() and
mthca_free() to use spin_lock_irqsave().
Signed-off-by: Roland Dreier <[email protected]>
Leave all SRQ methods out of the device's uverbs_cmd_mask if the
device doesn't have SRQ support (because of ancient firmware) so that
we don't allow userspace to call the driver's create_srq method. This
fixes a userspace-triggerable oops caused by ib_uverbs_create_srq()
following the device's ->create_srq function pointer, which will be
NULL if the device doesn't support SRQs.
Signed-off-by: Roland Dreier <[email protected]>
When destroying a QP, mthca locks both the QP's send CQ and receive
CQ. However, the following scenario is perfectly valid:
QP_a: send_cq == CQ_x, recv_cq == CQ_y
QP_b: send_cq == CQ_y, recv_cq == CQ_x
The old mthca code simply locked send_cq and then recv_cq, which in
this case could lead to an AB-BA deadlock if QP_a and QP_b were
destroyed simultaneously.
We can fix this by changing the locking code to lock the CQ with the
lower CQ number first, which will create a consistent lock ordering.
Also, the second CQ is locked with spin_lock_nested() to tell lockdep
that we know what we're doing with the lock nesting.
This bug was found by lockdep.
Signed-off-by: Roland Dreier <[email protected]>
Define a constant MTHCA_ARRAY_MASK to replace repeated uses of
(PAGE_SIZE / sizeof (void *) - 1) in mthca array code.
Signed-off-by: Roland Dreier <[email protected]>
Read the max_cmds value from the response to the QUERY_FW command
before printing out the value, so that the real value goes into the
debug output.
Signed-off-by: Roland Dreier <[email protected]>
Lockdep warns because uverbs is trying to take uobj->mutex when it
already holds that lock. This is because there are really multiple
types of uobjs even though all of their locks are initialized in
common code.
Signed-off-by: Roland Dreier <[email protected]>
Store away the user handle passed in from userspace when creating an
SRQ, so that the kernel can return the correct handle when an SRQ
asynchronous event occurs. (A 0 was incorrectly stored as the user
handle as part of the changes in 9ead190b, "IB/uverbs: Don't serialize
with ib_uverbs_idr_mutex")
Signed-off-by: Roland Dreier <[email protected]>
In ib_uverbs_cleanup_ucontext(), when iterating through the lists of
objects, there's no reason to do list_del() to remove the objects,
since both the objects and the lists that contain them are about to be
freed anyway. Since list_del() is a moderately big inline function,
getting rid of this extra work saves quite a bit of .text:
add/remove: 0/0 grow/shrink: 1/2 up/down: 3/-217 (-214)
function old new delta
ib_uverbs_comp_handler 225 228 +3
ib_uverbs_async_handler 256 255 -1
ib_uverbs_close 905 689 -216
Signed-off-by: Roland Dreier <[email protected]>
Currently, all userspace verbs operations that call into the kernel
are serialized by ib_uverbs_idr_mutex. This can be a scalability
issue for some workloads, especially for devices driven by the ipath
driver, which needs to call into the kernel even for datapath
operations.
Fix this by adding reference counts to the userspace objects, and then
converting ib_uverbs_idr_mutex into a spinlock that only protects the
idrs long enough to take a reference on the object being looked up.
Because remove operations may fail, we have to do a slightly funky
two-step deletion, which is described in the comments at the top of
uverbs_cmd.c.
This also still leaves ib_uverbs_idr_lock as a single lock that is
possibly subject to contention. However, the lock hold time will only
be a single idr operation, so multiple threads should still be able to
make progress, even if ib_uverbs_idr_lock is being ping-ponged.
Surprisingly, these changes even shrink the object code:
add/remove: 23/5 grow/shrink: 4/21 up/down: 633/-693 (-60)
Signed-off-by: Roland Dreier <[email protected]>
Documentation/infiniband/core_locking.txt says:
All of the methods in struct ib_device exported by a low-level
driver must be fully reentrant. The low-level driver is required to
perform all synchronization necessary to maintain consistency, even
if multiple function calls using the same object are run
simultaneously.
However, mthca's modify_qp, modify_srq and resize_cq methods are
currently not reentrant. Add a mutex to the QP, SRQ and CQ structures
so that these calls can be properly serialized.
Signed-off-by: Roland Dreier <[email protected]>
Some error paths after the mthca_alloc_mailbox() call in mthca_modify_qp()
just do a "return -EINVAL" without freeing the mailbox. Convert these
returns to "goto out" to avoid leaking the mailbox storage.
Signed-off-by: Roland Dreier <[email protected]>
Factor out common code for adding a userspace object to an idr into a
function idr_add_uobj(). This shrinks both the source and object code:
add/remove: 1/0 grow/shrink: 0/6 up/down: 57/-220 (-163)
function old new delta
idr_add_uobj - 57 +57
ib_uverbs_create_ah 543 512 -31
ib_uverbs_create_srq 662 630 -32
ib_uverbs_reg_mr 737 699 -38
ib_uverbs_create_cq 639 600 -39
ib_uverbs_alloc_pd 485 446 -39
ib_uverbs_create_qp 1020 979 -41
Signed-off-by: Roland Dreier <[email protected]>
In error paths when destroying an object, uverbs should not decrement
associated objects' usecnt, since ib_dereg_mr(), ib_destroy_qp(),
etc. already do that.
Signed-off-by: Roland Dreier <[email protected]>
The comparisons of priv->tx_tail to ah->last_send in ipoib_free_ah()
and ipoib_post_receive() are slightly unsafe, because priv->tx_lock is
not held and hence a stale value of ah->last_send might be used, which
would lead to freeing an AH before the driver was really done with it.
The simple way to fix this is to the optimization of early free from
ipoib_free_ah() and unconditionally queue AHs for reaping, and then
take priv->tx_lock in __ipoib_reap_ah().
Signed-off-by: Roland Dreier <[email protected]>
It's perfectly valid for a connection to an SRP target to have a
request limit of 0, so get rid of the message about it, which can spam
kernel logs even with printk_ratelimit(). Keep a count of such events
in a "zero_req_lim" SCSI host attribute instead, so someone who cares
can look at the statistics.
Signed-off-by: Roland Dreier <[email protected]>
The kernel has had wait_for_completion_timeout() for a long time now.
mthca should use it to handle FW commands timing out, instead of
implementing the same thing in a much more complicated way by using
wait_for_completion() along with a timer that does complete().
Signed-off-by: Roland Dreier <[email protected]>
Create an SRP FMR pool on HCAs that support FMRs, and use FMRs to map
gather/scatter lists that have more than one entry into a single
memory region that appears virtually contiguous to the SRP target
(which is the RDMA initiator).
This patch bails out on FMR mapping for SCSI commands where the
gather/scatter list cannot be mapped into a single FMR because there
are sub-page-sized entries in middle of the list. An unaligned
start or end of the list is OK.
Based on a patch by Vu Pham <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
In ib_umem_release_on_close(), if the kmalloc() fails, then a
reference to current->mm will be leaked. Fix this by adding a mmput()
instead of just returning on kmalloc() failure.
Signed-off-by: Roland Dreier <[email protected]>
If a reconnection attempt fails, then SRP does two scsi_host_put()s.
This is a historical relic from an earlier version of the driver that
took a reference on the scsi_host before trying to reconnect, so get
rid of the extra scsi_host_put().
Signed-off-by: Roland Dreier <[email protected]>
Sending a DREQ may fail, for example because the remote target has
already broken the connection. If so, then SRP should not wait for
the disconnection to complete, because it never will.
Signed-off-by: Roland Dreier <[email protected]>
Setting fw_cmd_doorbell allows FW command to be queued using posted
writes instead of requiring polling on a "go" bit, so it should be a
performance boost. However, the option causes problems with at least
some device/firmware combinations, so set the default to 0 until we
understand what's going on better.
Signed-off-by: Roland Dreier <[email protected]>
The ipath driver's table of PCI IDs needs a { 0, } entry at the end.
This makes all of the device aliases visible to userspace so hotplug
loads the module for all supported devices. Without the patch,
modinfo ipath_core only shows:
alias: pci:v00001FC1d0000000Dsv*sd*bc*sc*i*
instead of the correct:
alias: pci:v00001FC1d00000010sv*sd*bc*sc*i*
alias: pci:v00001FC1d0000000Dsv*sd*bc*sc*i*
Signed-off-by: Roland Dreier <[email protected]>
Signed-off-by: Bryan O'Sullivan <[email protected]>
When deleting a child interface with a non-default P_Key via
/sys/class/net/ibX/delete_child, the interface must be freed with
free_netdev() (rather than kfree() on the private data).
Signed-off-by: Roland Dreier <[email protected]>
Fix races in in destroying various objects. If a destroy routine
waits for an object to become free by doing
wait_event(&obj->wait, !atomic_read(&obj->refcount));
/* now clean up and destroy the object */
and another place drops a reference to the object by doing
if (atomic_dec_and_test(&obj->refcount))
wake_up(&obj->wait);
then this is susceptible to a race where the wait_event() and final
freeing of the object occur between the atomic_dec_and_test() and the
wake_up(). And this is a use-after-free, since wake_up() will be
called on part of the already-freed object.
Fix this in mthca by replacing the atomic_t refcounts with plain old
integers protected by a spinlock. This makes it possible to do the
decrement of the reference count and the wake_up() so that it appears
as a single atomic operation to the code waiting on the wait queue.
While touching this code, also simplify mthca_cq_clean(): the CQ being
cleaned cannot go away, because it still has a QP attached to it. So
there's no reason to be paranoid and look up the CQ by number; it's
perfectly safe to use the pointer that the callers already have.
Signed-off-by: Roland Dreier <[email protected]>
If a SCSI abort completes, or the command completes successfully, then
the driver must remove the command from its queue of pending
commands. Similarly, if a device reset succeeds, then all commands
queued for the given device must be removed from the queue.
Signed-off-by: Roland Dreier <[email protected]>
GuidInfo records have 8 byte GUIDs in them, so an index should be
multiplied by 8 to get an offset. mthca_query_gid() was incorrectly
multiplying by 16.
Noticed by Leonid Keller <[email protected]>.
Signed-off-by: Roland Dreier <[email protected]>
Roland Dreier