addresses shall access invalid descriptor DMA addresses on PCIe
hardwares and then panicked the system.
To fix it set descriptor DMA addresses before enabling Tx and Rx
such that hardware can see valid descriptor DMA addresses. Also
set RL_EARLY_TX_THRESH before starting Tx and Rx.
Reported by: steve.tell AT crashmail DOT de
Tested by: steve.tell AT crashmail DOT de
Obtained from: NetBSD
MFC after: 1 week
- First off, device drivers really do need to know if they are allocating
MSI or MSI-X messages. MSI requires allocating powerof2() messages for
example where MSI-X does not. To address this, split out the MSI-X
support from pci_msi_count() and pci_alloc_msi() into new driver-visible
functions pci_msix_count() and pci_alloc_msix(). As a result,
pci_msi_count() now just returns a count of the max supported MSI
messages for the device, and pci_alloc_msi() only tries to allocate MSI
messages. To get a count of the max supported MSI-X messages, use
pci_msix_count(). To allocate MSI-X messages, use pci_alloc_msix().
pci_release_msi() still handles both MSI and MSI-X messages, however.
As a result of this change, drivers using the existing API will only
use MSI messages and will no longer try to use MSI-X messages.
- Because MSI-X allows for each message to have its own data and address
values (and thus does not require all of the messages to have their
MD vectors allocated as a group), some devices allow for "sparse" use
of MSI-X message slots. For example, if a device supports 8 messages
but the OS is only able to allocate 2 messages, the device may make the
best use of 2 IRQs if it enables the messages at slots 1 and 4 rather
than default of using the first N slots (or indicies) at 1 and 2. To
support this, add a new pci_remap_msix() function that a driver may call
after a successful pci_alloc_msix() (but before allocating any of the
SYS_RES_IRQ resources) to allow the allocated IRQ resources to be
assigned to different message indices. For example, from the earlier
example, after pci_alloc_msix() returned a value of 2, the driver would
call pci_remap_msix() passing in array of integers { 1, 4 } as the
new message indices to use. The rid's for the SYS_RES_IRQ resources
will always match the message indices. Thus, after the call to
pci_remap_msix() the driver would be able to access the first message
in slot 1 at SYS_RES_IRQ rid 1, and the second message at slot 4 at
SYS_RES_IRQ rid 4. Note that the message slots/indices are 1-based
rather than 0-based so that they will always correspond to the rid
values (SYS_RES_IRQ rid 0 is reserved for the legacy INTx interrupt).
To support this API, a new PCIB_REMAP_MSIX() method was added to the
pcib interface to change the message index for a single IRQ.
Tested by: scottl
control data but no payload data is passed.
Change m_uiotombuf() to return at least one empty mbuf if the requested
length was zero. Add comment to sosend_dgram and sosend_generic().
Diagnoses by: jhb
Regression test by: rwatson
Pointy hat to. andre
--------------------------
[Deadlock] is caused by a lock order reversal in vfs_lookup(), where
[some] process is trying to lock a directory vnode, that is the parent
directory of covered vnode) while holding an exclusive vnode lock on
covering vnode.
A simplified scenario:
root fs var fs
/ A / (/var) D
/var B /log (/var/log) E
vfs lock C vfs lock F
Within each file system, the lock order is clear: C->A->B and F->D->E
When traversing across mounts, the system can choose between two lock orders,
but everything must then follow that lock order:
L1: C->A->B
|
+->F->D->E
L2: F->D->E
|
+->C->A->B
The lookup() process for namei("/var") mixes those two lock orders:
VOP_LOOKUP() obtains B while A is held
vfs_busy() obtains a shared lock on F while A and B are held (follows L1,
violates L2)
vput() releases lock on B
VOP_UNLOCK() releases lock on A
VFS_ROOT() obtains lock on D while shared lock on F is held
vfs_unbusy() releases shared lock on F
vn_lock() obtains lock on A while D is held (violates L1, follows L2)
dounmount() follows L1 (B is locked while F is drained).
Without unmount activity, vfs_busy() will always succeed without blocking
and the deadlock isn't triggered (the system behaves as if L2 is followed).
With unmount, you can get 4 processes in a deadlock:
p1: holds D, want A (in lookup())
p2: holds shared lock on F, want D (in VFS_ROOT())
p3: holds B, want drain lock on F (in dounmount())
p4: holds A, want B (in VOP_LOOKUP())
You can have more than one instance of p2.
The reversal was introduced in revision 1.81 of src/sys/kern/vfs_lookup.c and
MFCed to revision 1.80.2.1, probably to avoid a cascade of vnode locks when nfs
servers are dead (VFS_ROOT() just hangs) spreading to the root fs root vnode.
- Tor Egge
To fix the LOR, ups@ noted that when crossing the mount point, ni_dvp
is actually not used by the callers of namei. Thus, placeholder deadfs
vnode vp_crossmp is introduced that is filled into ni_dvp.
Idea by: ups
Reviewed by: tegge, ups, jeff, rwatson (mac interaction)
Tested by: Peter Holm
MFC after: 2 weeks
sparc64 GENERIC and the sound device drivers known working on sparc64
to use bus_get_dma_tag() to obtain the parent DMA tag so we can get rid
of the sparc64_root_dma_tag kludge eventually. Except for ath(4), sk(4),
stge(4) and ti(4) these changes are runtime tested (unless I booted up
the wrong kernels again...).
a power saving mode otherwise.
- If the thread is already bound in sched_bind() unbind it before
re-binding it to a new cpu. I don't like these semantics but they are
expected by some code in the tree. Patch by jkoshy.
Dont expose em->shared to the outside world before its properly
initialized. Might not affect anything but its at least a better
coding style.
Dont expose em via p->p_emuldata until its properly initialized.
This also enables us to get rid of some locking and simplify the
code because we are workin on a local copy.
In linux_fork and linux_vfork create the process in stopped state
to be sure that the new process runs with fully initialized emuldata
structure [1]. Also fix the vfork (both in linux_clone and linux_vfork)
race that could result in never woken up process [2].
Reported by: Scot Hetzel [1]
Suggested by: jhb [2]
Reviewed by: jhb (at least some important parts)
Submitted by: rdivacky
Tested by: Scot Hetzel (on amd64)
Change 2 comments (in the new code) to comply to style(9).
Suggested by: jhb
work when we start requiring this.
- Don't specify an alignment when creating our own parent DMA tag;
the supported DMA engines require no alignment constraint (f.e. the
LANCE child does though) and it's no inherited by the child DMA
tags anyway (which probably is a bug though).
- Fix whitespace nits.
These are shared-memory variants based on Am79C90-compatible chips
that apart from the missing DMA engine are similar to the 'ledma'
variant including using a (pseudo-)bus/device for the buffer that
the actual LANCE device hangs off from. The performance of these is
close to that of the 'ledma' one, like expected at a few times the
CPU load though.
1) Do not do quota accounting for the actual quota data files
or for file system snapshot files ("system" files). This
prevents a deadlock descibed in PR kern/30958 if the kernel
ever has to grow the quota file. Snapshot files were already
exempt from the quota checks, but this change generalized the check.
2) Fix a cast that caused extremely large uids/gids to incorrectly
write the quota information to the data file at a truncated
value for a uint_t32 id value. The incorrect cast caused quota
files in this case to be around 4GB in size, with the correct cast
they can now be 131GB in size. Also related to PR kern/30958.
3) Check for what appear to be negative UIDs/GIDs and not account
for them. This prevents the quota files from becoming 131GB in
size and causing quotacheck to run forever at bootup. This could
also cause the kernel to try and expand the quota file, which might
deadlock due to the issue in #1. kern/30958 and kern/38156
(and some much older closed PR's).
4) With the deadlock problems gone, the kernel can now expand the
size of the quota database files if it needs to.
5) Pass in the i-node count change value to chkiq and chkiqchg as an
int, like it used to be before the common routine was split up
into 2 different routines to increase / decrease the i-node in-use
count. Prevents an underflow on the i-node count. Related
to PR kern/89247.
6) Prevent the block usage from growing slowly if a file system is
full and the write was denied due to that fact. PR kern/89247.
Some of these changes require an updated quotacheck to prevent
the creation of huge (131GB) quota data files (item #3).
#1/#4 probably fixes a lot of the random hangs when quotas are enabled,
possibly some of the jail hangs.
unlike documented may not take effect without an initialization. So
don't invoke (*sc_mediachange) directly in lance_mediachange() but
go through lance_init_locked(). It's suboptimal to impose this for
all chips but given that besides the affected PCI bus front-end the
only other front-end which supports media selection is and likely
ever will be the 'ledma' front-end I see not enough reason to break
the in-driver API for this (though one could argue both ways here).
the ipi settings. If NEEDRESCHED is set and an ipi is later delivered
it will clear it rather than cause extra context switches. However, if
we miss setting it we can have terrible latency.
- In sched_bind() correctly implement bind. Also be slightly more
tolerant of code which calls bind multiple times. However, we don't
change binding if another call is made with a different cpu. This
does not presently work with hwpmc which I believe should be changed.
front of isp_init so we can read NVRAM even if we're role ISP_NONE.
Prepare for reintroduction of channels (for FC) for N-Port
Virtualization.
Fix a botch in handle assignment that caused us to nuke one device
when a new one arrives and end up with two devices with the same
identity in the virtual target mapping table.
ifmedia_init() invocation. IFM_IMASK makes only sense here when all of
the maxium of 32 PHYs on each one MII bus support disjoint sets of media,
which generally isn't the case (though it would be nice if we had a way
to let NIC drivers indicate that for the few card models where the PHY
configuration is known/fixed and IFM_IMASK actually makes sense).
- Add and use a miibus_print_child() for the bus_print_child method which
additionally prints the PHY number (which actually is the PHY address)
so one can figure out the media instance <-> PHY number mapping from the
PHY driver attach output. This is intented to be usefull in situations
where the addresses of the PHYs on the bus are known (f.e. of internal/
integrated PHYs) so one can feed the appropriate media instance number
to ifconfig(8) (with the upcoming change for ifconfig(8)).
This is more or less inspired by the NetBSD mii_print().
multiple PHYs. In case some PHYs currently driven by ukphy(4) exhibit
problems when isolating due to incomplete implementations or silicon bugs
we'll need to add specific drivers for these. Looking at NetBSD and
OpenBSD I don't expect problems here though (quite the contrary; we still
seem to set MIIF_NOISOLATE without good reason in a bunch of PHY drivers).
- Fix a style(9) whitespace nit.
capability rather than hardcoded offsets for a particular card. While
I'm here, expand the constants some.
- Change the ahd(4) driver to use pci_find_extcap() to locate the PCI-X
capability to keep up with the first change.
Reviewed by: scottl, gibbs (earlier version)
- Switch back to direct modification of remote CPU run queues. This added
a lot of complexity with questionable gain. It's easy enough to
reimplement if it's shown to help on huge machines.
- Re-implement the old tdq_transfer() call as tdq_pickidle(). Change
sched_add() so we have selectable cpu choosers and simplify the logic
a bit here.
- Implement tdq_pickpri() as the new default cpu chooser. This algorithm
is similar to Solaris in that it tries to always run the threads with
the best priorities. It is actually slightly more complex than
solaris's algorithm because we also tend to favor the local cpu over
other cpus which has a boost in latency but also potentially enables
cache sharing between the waking thread and the woken thread.
- Add a bunch of tunables that can be used to measure effects of different
load balancing strategies. Most of these will go away once the
algorithm is more definite.
- Add a new mechanism to steal threads from busy cpus when we idle. This
is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The
threshold is the required length of a tdq's run queue before another
cpu will be able to steal runnable threads. This prevents most queue
imbalances that contribute the long latencies.
headers in .S directly rather than getting to their macros through
genassym.c/assym.s so there are less headers genassym.c has to be
kept in sync with.
While at it fix some stytle(9) bugs (indentation, prototype format,
sort headers, etc) and remove trailing whitespace.
that can be used to check whether receive data is ready, i.e. whether
the subsequent call of uart_poll() should return a char, and unlike
uart_poll() doesn't actually receive data.
- Remove the device-specific implementations of uart_poll() and implement
uart_poll() in terms of uart_getc() and the newly added uart_rxready()
in order to minimize code duplication.
- In sunkbd(4) take advantage of uart_rxready() and use it to implement
the polled mode part of sunkbd_check() so we don't need to buffer a
potentially read char in the softc.
- Fix some mis-indentation in sunkbd_read_char().
Discussed with: marcel
may also reflect a Fireplane/Safari or JBus bus (or a virtual bus which
in turn reflects a JBus bus or something like that...).
- In the both the sparc64 and sun4v bus_machdep.c use __FBSDID.
- Spell SBus the official way in comments.
- Replace hardcoded function names (all of which were actually outdated)
in panic and status strings with __func__.
- Fix whitespace nits.
hooks get their per hook rcvdata methods, and all functions are organized
corresponding to protocol stack model.
Submitted by: Alexander Motin <mav alkar.net>
Reviewed by: archie, julian
and friends along with all hacks required to implement them. None of
the drivers currently built (as part of GENERIC, LINT or modules) on
sparc64 or sun4v and none of those we might want to use there in
future uses them, AFAICT there actually never was a driver hooked up
to the sparc64 or sun4v build that correctly used these functions
(and it looks like that due to a bug read{b,w,l}()/write{b,w,l}() and
the other functions working on a memory handle never actually worked on
sun4v). All they ever were good for on sparc64 and sun4v was erroneously
dragging in dependencies on isa(4) in drivers like f.e. dpt(4), si(4)
and syscons(4) in source files that supposedly were bus-neutral and
hiding issues with drivers like f.e. ng_bt3c(4) that used these
functions with busses other than isa(4) and therefore couldn't work on
these platforms.
the newly added DEV_EISA. This is done so that these back-ends can
be compiled on platforms not providing in{b,w,l}()/out{b,w,l}() and
friends (but may wish to use them together with bus front-ends other
than the EISA one).
- Finally all splxx() are removed
- Count error fixed in mapping array which might
cause a wrong cumack generation.
- Invariants around panic for case D + printf when no invariants.
- one-to-one model race condition fixed by using
a pre-formed connection and then completing the
work so accept won't happen on a non-formed
association.
- Some additional paranoia checks in sctp_output.
- Locks that were missing in the accept code.
Approved by: gnn
to open() [1].
Improve locking for accessing session control structures [2].
Try to document (most likely harmless) races in the code [3].
Based on submission by: Intron (intron at intron ac) [1]
Reviewed by: jhb [2]
Discussed with: netchild, rwatson, jhb [3]
