This commit introduces support for the HMB feature and will setup the
buffer during start-up. The host-memory-buffer (HMB) feature is mostly
used on NVMe devices that do not make use of an DRAM cache to store its
translation tables amongst other operational data. Not using HMB can
impair the performance on such devices.
The memory is allocated in 2 MiB chunks of DMA-capable memory and its
total size in bytes is configurable via the 'hmb_size' config attribute.
The driver always checks the minimal and preferred size of the HMB and
issues a warning in case it is not enabled via the configuration.
Moreover, if the configured size is less than the minimal amount
required by the device the HMB is not configured at all and a warning
is issued also. If the configured size is more than the preferred size
it will be capped to that amount.
Fixes#4715.
This patch converts the loader implementation to the 'Session_object'
interface to implement the ability to upgrade the session caps and RAM
after after the session creation. This is needed to accommodate the
change of the Loader::Connection in commit "Remove format strings from
connection types".
Issue #2064
This patch ultimatedly removes format strings from Genode's base API.
Users of the former base/snprintf.h and base/console.h headers may
use the free-standing 'format' library hosted in the ports repository.
Fixes#2064Fixes#3869
This little library implements rudimentary format-string support. It is
useful for porting 3rd-party code that ought not depend on a full libc.
Issue #2064
The new 'Connection' constructor accepts the session label, affinity,
and args as constructor arguments. The session arguments are passed as a
'Genode::String'. This allows for side-stepping the need for rendering a
format string passed to 'Env::session'.
Issue #2064
The additional files provide Chacha implementation in case
may_use_simd() returns false because in_interrupt() returns true.
Related to issue #4778, commit "lx_emul: track irq state changes"
Implement udelay for all architectures within common lx_emul code,
remove driver-specific version.
Always do busy looping in udelay and use cpu_relax()
like the original Linux code. Thereby, we profit from architecture specific
instructions (cpu cool down), and jiffies are updated if irqs are on.
Ref genodelabs/genode#4778
There are rare use-cases where cpu_relax is used inside a busy loop
(i2c bus functions), which only will break when the jiffies counter
reaches a specific value.
Because of the cooperative scheduling done in lx_emul, no timer
interrupt will break such a loop. As a workaround, we check for
necessary jiffies updates inside cpu_relax if interrupts are enabled.
Ref genodelabs/genode#4778
To be able to check for interrupts being on or off,
the enabling and disabling is tracked in lx_emul.
When interrupts get received, they have to be turned off.
Ref genodelabs/genode#4778
Do not start and stop idle ticking within the timer interrupt
routine, but do it around the whole Lx_kit scheduling, which
is always called when Linux code gets active again, either
because of backend signals, interrupts, or timing signals.
This commit implicitly reverts the (incomplete) solution of
issue #4550
Ref genodelabs/genode#4778
Linux kernel static functions usb_string_sub() and usb_get_langid() were
made accessible to implement robust string rerieval.
Fixes#4756Fixes#4757Fixes#4772
The namespace draws a clear line between the base library and the core
component.
It is declared at the new core-local header <types.h>, which is expected
to be included by all code of the core component. It is thereby a
natural place for kernel-agnostic general types like commonly used C++
utilities.
Fixes#4777
Implicitely fixes problems with USB devices having more than 4G blocks.
Formerly the 16-Cmd LBA requests were silently casted to 32-bit.
Fixgenodelabs/genode#4771
In case the packet is erronous the value of 'actual_size' can be
invalid and using it may lead to a page-fault due to out-of-bounce
access.
With this commit access is only performed on successful packets.
Fixes#4763.
Workers are only summoned by the manager or when there are no workers on
a work queue. In case a work in front of a work queue depends on a work
behind it, a deadlock may occur. To solve this Linux spawns a rescue
worker using timers. Timing of Linux based drivers is currently reworked
and not reliable. Therefore, we increase the number of workers that are
spawned from 1 to 3 per work queue in order to resolve possible
deadlocks.
Fixes#4762
Set wakeup pointer in FADT/FACS tables and prepare/place ACPI resume code
at application processors (AP) 16-bit entry. Exclude memory used for AP resume
from RAM range to avoid usage by Genode core in allocators.
Issue #4669
The CPUs are woken all at once and up to now the IDs are assigned depending
on the arrival order, which is unfortunate for the resume case. Keep track
of once assigned IDs for CPUs, so that on resume the very same CPU id is given.
The APIC id is guaranteed to be fixed per CPU on suspend, but unfortunately
not dense packed.
Issue #4669
The ACPI table FACS and FADT are required to support ACPI suspend/resume. The
commits add the lookup of the ACPI table in bootstrap and the general usage
of the ACPI registers via the MMIO framework.
Issue #4669
Add explicit init() to Timer infrastructure to re-initialize the hardware
based on the parameters given during constructing time of the timer object.
Issue #4669
On resume the CPU state must be re-initialized, since all CPUs have been
potentially off (depending on sleep form) and lost state. Handle the halt
and suspend state explicitly by an extra Job implementation which can be
executed without holding the global kernel lock.
Issue #4669
Instead of re-using the register values found in HCR_EL2 and CPTR_EL2
and setting single bits within them, define the complete content to
prevent inconsistent hardware/hypervisor state.
Ref genodelabs/genode#4759
Several explicit casts could not be avoided yet, due to the missing
differentiation in between virtual and physical addresses that leads
to casting problems when using 32-bit ARM, and because the MMIO
register framework does not allow to return narrowed types of bitfields.
Apart from that, this commit fixes a switch-case fallthrough error in
Mmio_register::write.
Fixgenodelabs/genode#4770
The read_config and write_config functions in the generic virtio
headers used by all drivers lead to compiler warnings resp. errors
if effective-c++ switch is enabled. Moreover, the functions require
to define the access width as parameter. We can better turn them
into template functions using the value type to read resp. write to
derive the access width.
Ref genodelabs/genode#4344
The 'file_size' type denotes the size of files on disk in bytes. On
32-bit architectures it is larger than the size_t, which refers to
in-memory object sizes.
Whereas the use of 'file_size' is appropriate for ftruncate and seek, it
is not a suitable type for the parameters of read/write operations
because those operations refer to in-memory buffers.
This patch replaces the use of 'file_size' by size_t. However, since it
affects all sites where the read/write interface is uses, it takes the
opportunity to replace the C-style (pointer, size) arguments by
'Byte_range_ptr' and 'Const_byte_range_ptr'.
Issue #4706
Prior to this change the libdrm Lima implementation supported the
creation of multiple contexts where each context, however, was
treated as the same client like it was done in the Lima driver
itself.
With this commit each context becomes its own client while the
main context always performs all buffer object related allocation
and the other context import each needed BO before submitting.
Fixes#4760.
Sync file-descriptors a purely virtual and are not known by the libc and
should be only accessed by 'drm_poll'. To prevent accidental access move
them to range never used by the libc fd allocator.
Issue #4760.
We determine the owner of a Vram object by the session cap of the
session that created the object. We should not copy this cap to other
places because this increases reference counting, that can become very
large with many objects. Therefore, we pass a wrapper 'Owner' object
containing the cap by reference.
issue #4713
* 'map_buffer', GTT mappings are currently not used by Iris
* 'set_tiling', tiling is currently not used by Iris because there are
no GTT mappings
issue #4713
Change the abstraction from buffers to video RAM (VRAM). The notion of
buffers can be provided at the client side (e.g., Mesa) and multiple
buffers can be there be associated to one VRAM area, thus saving
resources (meta data overhead) when allocating many buffers. A VRAM area
can also be mapped to one single buffer as before for clients or drivers
that do not take advantage of this feature.
issue #4713
by checking for it in core, submitting a signal (if registered) and not
replying instantly. Before this commit, an exception (a non page fault)
caused and endless "unknown exception" loop between originator and core.
This was handled before commit "core: kernel-agnostic 'Mapping' type" by
throwing an exception, which was with that commit not working anymore.
Fixes#4751
Segment assignment stays valid for all sections following .dynamic,
which implies that .got and co also end up in the PT_DYNAMIC segment.
This is not intended. Therefore, we move the .dynamic section before
bss and assign bss to the rw PT_LOAD segment only.
fixes#4750
The screenshot trigger displays a little red dot at the upper-left
corner of the screen. When touched or clicked-on, it generates an
artificial key-press-release sequence for the print key and disappears
for one second. In this time, a separate screenshot component can handle
the print key by capturing the screen without the red dot appearing in
the saved picture.
The touch-keyboard config accepts the new attributes 'opaque="yes"
and 'background=#112233' to control the dialog background. The
attributes are passed unmodified to embedded the menu view.
This patch add the support needed for implementing on-target
system-update functionality. The new 'empty' method is required for the
two-step procedure of updating image-index files - removing the old
index, followed by downloading the most current one. It is used to the
detect the completion of the first step.
Issue #4744
This patch reflects more details such as the percentage value for
downloads maintained in the download queue. This makes this information
easier to display at the GUI.
Issue #4744
This patch replaces the generic "In progress..." message by slightly
more concrete information about the type of operation. This is useful to
distinguish the two steps during the expansion of a partition.
This patch enhances the depot_download subsystem with support for
downloading and querying system images.
The installation ROM support the following two now download types:
<image_index path="<user>/image/index"/>
<image path="<user>/image/<name>"/>
Internally, the depot-download subsystem employs the depot-query
component to determine the missing depot content. This component
accepts the following two new queries:
<images user="..."/>
<image_index user="..."/>
If present in the query, depot_query generates reports labeled as
"images" and "image_index" respectively.
The also tracks the completion of each job depending on the depot-
query results, so that the final report contains a result for each
installation item requested. Prior this patch, the inactivity of the
depot-download manager (indicated by an empty state report) was
interpreted as success. But that prevents the proper association of
results and requested installation items.
Issue #4744
The 'scan' query for depot users used to list the names of depot users.
However, the URL and existence of a pubkey is useful to know when
assembling a GUI menu from this list.
Issue #4744
This patch improves the Readonly_file::read method such that the
capacity of the specified buffer is used as upper bound for the read
operation instead of VFS-internal I/O buffer sizes. This relieves the
caller from implementing a read loop in most cases.
As a step away from C-ish use of the API, the patch deprecates the old
'read' method that takes the buffer as char *, size_t arguments.
Fixes#4745
Each time when enlarging the menu view, a new 'Gui_buffer' is
constructed with the new size. At its contruction time, the default
reset background color is in effect, which is evaluated by
'reset_surface' as part of the construction. A custom reset color
takes effect only after the construction. Hence, when the Gui_buffer is
flushed to screen immediately at construction time, the gray default
becomes visible for a short time.
This patch changes the Gui_buffer to accept the reset background color
as construction argument so that it takes immediate effect.
Related to #4592
The new utility returns a key code for a passed name and is implemented
by linear search, which is slow but sufficient in situations like config
updates.
Issue #4748
The leading newline in the _capture function confused _filecontent,
which resulted in empty port hashes.
Fixes regression introduced in 437fd21ba0
(issue #4725).
lwip uses a sophisticated reference-counting scheme in chains of pbufs,
which the former manual implementation of read() for TCP data broke.
Using pbuf_free_header() keeps the chain intact and also relieves our
implementation from the burden of "offset" maintenance.
Fixes#4722
Due to lacking hardware access the driver was test with one RTL8188EE
based device, namely [10ec:8179] (rev 01), only. As the access to the
PCI config space is restricted the driver loads the non power-saving
FW and the driver port is therefor only tested with that. The
accesses are documented should we choose to support them one way or
another later on.
The 'wifi.run' run script as well as Sculpt served as testing ground
where the driver worked fine so far.
Fixes#4714.
Check if we are able to submit the Usb packet before attempting to
allocate it. In case the queue is already full we wait until another
pending packet is freed.
Fixes#4701.
Up to now it was only checked if an issued admin command was processed
in a timely fashion. Otherwise it has been treated as failed.
However, the completion-queue entry was not examined and the caller was
not able to access the entry itself. Depending on the command, checking
the completion-queue entry might be necessary, e.g. GET/SET_FEATURE.
Issue #4715.
Since the 'Platform::Device' constructor will defer the creation until
the content of the devices ROM is valid performing the PRP list helper
creation afterwards should be done with valid IOMMU information.
Issue #4715.
Some USB devices (e.g., webcams) fail to deliver their configuration
descriptor early after power-up. Testing revealed that retrying the
requests usually succeeds on second or third attempt.
Fixes#4739
Genode build system allows to easily enable 'ccache' for builds. This
change allows to enable using 'ccache' also for build of reference
Linux kernel used during porting device drivers.
To enable 'ccache' it is enough to pass value of 'CC' variable when
executing Linux build but this build by default depends on time when it
is built which causes 'ccache' misses. To solve this issue additional
flags are passed to make build independent from time, current user and
host on which build is performed.
Issue #4718
With the new 'presets:' tag, .sculpt files can now refer to deploy
configurations to be integrated in the presets/ subdirectory of the
config file system. Those files can thereby be used as preconfigured
system scenarios. Such a preconfigured scenario can be loaded at
runtime by copying the preset file to config/deploy.
Issue #4731
In case of very rapid config ROM change (<300ms), the linux driver may still
be in progress of applying the previous change (e.g. switching connector on/off).
During this progress all tasks may become not runnable (waiting for IRQ/timeouts
until hardware state settles), the newest config ROM change/signal gets
dispatched, which lead to continuing the previous change request, but not
to re-starting/re-applying the new config change. To avoid this situation,
explicitly track whether a previous config change was finished and track if an
interim config change request came in. If so, re-start the lx_user
task with the newest config change.
Fixes#4721
* Update links from forward rules only with forward rules and links from
transport-routing rules only with transport-routing rules. Besides raising
the performance of the code, this also fixes a former bug that allowed
forward-rule links to falsely stay active because of a transport-routing
rule that matched the client destination ip and port.
* Don't use good-case exceptions for updating TCP/UDP links on re-configuration
of the router.
* Make conditions when to dismiss a forward rule easier to read.
* Introduces != operator to the public Port class in the net library.
* Fix unnecessary log message that a link was dismissed when only a potentially
matching forward rule turned out to be not matching.
* Apply Genode coding style to if statements with a single body statement.
Fix#4728
This fixes a bug that was introduced by this earlier commit:
"nic_router: find forward rules w/o exceptions"
The NIC router used to falsely dissolve TCP/UDP connection states when
reconfiguring although the connection states were still legal according to the
new config. The reason was that the above mention commit nested lambdas but
missed to return from the last nesting level when having found a configuration
that legitimates the connection state.
Ref #4728
Tests on qemu would fail when started with RAM sizes from 1025MiB to
2048MiB, because the the mapping hole in the page table from 1GiB to
2GiB would interfere with qemu's mapping addresses for ACPI.
Identity-map the complete first 4GiB of memory to catch all early
memory accesses during bootstrap.
Fixes#4724.
This patch simplifies the 'Deploy::update_managed_deploy_config'
interface by keeping an internal copy of the currently used deploy
template inside the 'Deploy' class. The template is updated whenever
the config/deploy file is modified.
This change weakens the coupling between the '_manual_deploy_rom' and
the '_deploy' subsystem, easing the upcoming implementation of the
switching between presets.
Adds befriended test-local wrappers for the classes Cpu_share and Cpu_scheduler
and adds a print method to the scheduler wrapper that prints the internal state
of the scheduler to the given output. Cpu_shares are referenced in the output
via a the IDs that the test uses to organize them. I.e., this corresponds to
how the CPU shares are named when calling the atomic steps the test is made of.
Ref #4151
Ref #4710
This adapts the test to the changes that were applied to the scheduling scheme
by the following commits:
* base-hw scheduler: optimize quota depletion events
* base-hw scheduler: fix bug on removing head
* base-hw scheduler: fix ready method
* base-hw: optimize & cleanup scheduler
Part of that is that the test used to check whether the act of setting a share
ready outdates the head or not. However, with the current version of the
scheduler, this check is not possible anymore. We can merely check whether the
head is outdated after setting the share ready. So, among other adaptions, this
commit adapts the expectations of the test to the new semantics of the check.
Ref #4151
Ref #4710
* Get rid of preprocessor macros.
* Introduce Main as class.
* Exit with -1 instead of endless loops on errors.
* Don't try to deal with error conditions, just print a message and exit
with -1.
* Only one operation per line.
Ref #4151
Ref #4710
This is an optimization for the case that a prioritized scheduling context
needs slightly more time during a round than granted via quota. If this is the
case, we move the scheduling context to the front of the unprioritized schedule
once its quota gets depleted and thereby at least ensure that it does not have
to wait for all unprioritized scheduling contexts as well before being
scheduled again.
Note that this introduces the possibility of undeserved starvation of
unprioritized scheduling contexts to the scheduling scheme. If there are
enough prioritized contexts that deplete their quota during a round,
they may cover up also the rest of the round with their unprioritized time
slices. If this happens every round, contexts without a priority/quota may
never get a turn. In the previous scheduling scheme, this could not occur as
the unprioritized schedule was completely independent from prioritized
schedules and rounds.
Ref #4151
Ref #4710
The scheduler did not consider the consumed quota during a call to "update"
if the head that consumed the quota was removed from the scheduler. When this
occured, the internal round time did not advance as expected but remained at
its previous value untile the next call to "update" (without a removed head)
This commit introduces a new flag that is set only when the head gets removed
in order to detect and handle the situation correctly on the next call to
"update".
Ref #4151
Ref #4710
Setting the _need_to_schedule member in the 'ready' method of the scheduler
was not done correctly. At least, the _need_to_schedule was set true in
situations were the head was not outdated by the 'ready' operation.
Ref #4151
* Remove *request* in context of: wait, reply, send to shorten it.
* Use ready_to_* instead of can_*, which is regularily used in Genode's APIs
* Replace helping_sink with helping_destination, as destination is more common
Ref genodelabs/genode#4704
The IPC protcol violations are:
* Sending to an unknown thread (cap)
* Waiting for messages if a reply hasn't happened yet
This silents threads that otherwise repeatedly cause kernel messages
about the violation.
Ref genodelabs/genode#4704
* Split the internal state into incoming and outgoing message relations
* Avoid fragmenting of one state like formerly '_state' and '_help'
* Remove pointer to caller, use incoming FIFO instead
This commit fixes at least two bugs that were triggered by tests that
destroy threads in many different states, like run/bomb:
* The '_help' data member was not reset reliable in each situation where a
helping relationship came to an end. However, when we fixed this bug alone
in the old state model, the issues remained. The new state model fixes
this bug as well.
* A thread sometimes referenced an already dead thread as receiver. This caused
the kernel IPC code to access the vtable of an object that didn't exist any
longer. Note that the two threads were not in direct IPC relationship while
the receiver was destroyed, so, there must have been an intermediate node
between them. Due to the complexity of this problem, we eventually gave up
pin-pointing the exact reason in the kernel IPC code. The issue disappeared
with the new state model.
Fixgenodelabs/genode#4704
When writing the GPT header, the tool always wrote the GPT entries
belonging to the primary header to LBA following the header. Normally
this is LBA 2 as the header is located in LBA 1. The GPT allows for
up to 128 entries that all in all cover 16 KiB of storage space.
However, on some systems, e.g. ARM-based machines, the bootloader can
be stored in this region. For this reason the GPT entries may be moved
to a different LBA.
This commit changes the tool to adhere to then given GPE LBA in header
when writing out the modified GPT data.
Fixes#4720.
The old 'Io_response_handler::io_progress_response' interface has been
replaced by the 'Vfs::Env::User::wakeup_vfs_user' (issue #4697). The
remaining 'read_ready_response' method is now hosted in the
appropriately named 'Read_ready_response_handler'.
Issue #4706
This patch keeps driving the internal state machines until no progress
can be made. This required fixing the return values of several execute
functions, which used to report progress while being in complete state.
Along the way, the patch removes default switch cases to ensure that all
states are covered.
Issue #4706
This commit supplements the various I/O signal handlers of the VFS
plugins with calls of the new 'Vfs::Env::User::wakeup_vfs_user'
interface, which will subsequently replace the old 'Io_progress_handler'
(issue #4697).
Issue #4706
The 'blocked_handles' queue was used to notify the VFS user via the
'io_progress_response' mechanism. This is now covered by the
'wakeup_vfs_user' interface introduced in issue #4697.
Issue #4706
Information about PS/2 and PIT where moved to app/pci_decode in the
following commit.
pci_decode: report devices from ACPI info
We still provide an empty <devices> node as the file itself is used by
platform agnostic run scripts.
When running on x86, and riscv never enter the kernel for cache maintainance,
but use the dummy implementation of the generic base library instead.
On ARMv8 it is not necessary to enter privileged mode for cache cleaning, and
unification of instruction/data cache, but only for invalidating cache lines
at all levels, which is necessary for the use cases, where this function it
needed (coherency of DMA memory).
Fixgenodelabs/genode#4339
This call is used to query the cache line size of the underlying CPU.
For now it is only implemented and used by 'arm_v8' platforms.
It does not distinguish between D-/I-cache sizes and always uses the
smallest size. Furthermore it does not account for any discrepancy
in 'big.little' CPUs.
Issue #4339.
To prevent the kernel to deadlock, or call itself with a syscall when
using a lock potentially hold by a core thread, the log console's
backend for core (hw) gets replaced by a specific variant that checks
whether it runs in the kernel context before using the mutex.
Fixgenodelabs/genode#3280
When a domain receives a new dynamic router IP address and that domain has
active connection states (TCP/UDP/ICMP) from another domain with NAT applied,
the connection states used to stay active while becoming obsolete. They
become obsolete because their identification and their packet processor
use the old routers IP address due to NAT.
One consequence was that connections became dysfunctional when the server
domain received a new dynamic router IP address. Request packets were still
routed from client to server, but when entering the server, their source IP
address was the outdated router address. Consequently, the server responses
used the outdated address as destination and the router dropped the responses
because it did not know this address anymore.
This commit fixes the problem by letting a domain destroy all its connection
states that were initiated from within other domains whenever it detaches from
its current IP configuration.
Strictly speaking, it is not necessary to destroy all connection states, only
those that the domain applies NAT to. However, the Genode AVL tree is not built
for removing a selection of nodes and trying to do it anyways is complicated.
So, for now, we simply destroy all connection states.
Note that the other way around was handled correctly already. When a domain
detaches from its IP config, all interfaces of that domain destroy all the
connection states they created (towards other domains).
Fixes#4696
If the IP config does not change on updates to the router IP config of a domain
change (a common case on DHCP RENEW), prevent detaching from the old config and
attaching to the new one. Because this would not only create unnecessary CPU
overhead but also force all clients at all interfaces that are listening to
this config (via config attribute 'dns_config_from') to restart their
networking (re-do DHCP).
Ref #4696
Check 'pv == nullptr' in 'ShClSvcImplWriteData' and return
VERR_INVALID_POINTER if invalid (as is done, for example, in the X11
implementation).
issue #4666
By adding a 'write_ready' interface following the lines of the existing
'read_ready', VFS plugins become able to propagate the (de-)saturation
of I/O buffers to the VFS user. This information is important when using
a non-blocking file descriptor for writing into a TCP socket. Once the
application observes EAGAIN, it expects a subsequent 'select' call to
return as soon as new I/O buffer space becomes available.
Before this patch, the select call would always return under this
condition, causing an unnecessarily busy write loop.
Issue #4697
The new interface is meant to replace the 'Vfs::Io_response_handler'.
In contrast to the 'Io_response_handler', which had to be called
on a 'Vfs_handle', the new interface does not require any specific
'Vfs_handle'. It is merely meant to prompt the VFS user (like the libc)
to re-attempt stalled I/O operations but it does not provide any
immediate hint, about which of the handles have become ready for
reading/writing.
Issue #4697
This patch removes the 'Insufficient_buffer' exception by returning the
WRITE_ERR_WOULD_BLOCK result value instead. It also eliminates the
superfluous WRITE_ERR_AGAIN and WRITE_ERR_INTERRUPT codes.
Issue #4697
The original size of 16K impedes the batched processing of network
packets. Changing the value to 256K reduces the number of context
switches when downloading large files and thereby improves the
throughput by more than 25% (base-hw on qemu_x86_64, using fetchurl
to download a file of 100 MiB via the NIC router from lighttpd).
Issue #4697
This patch fosters the batching of network packets transferred by the
lwIP stack over the NIC connection. It replaces the eager submission of
the packet-stream's data-flow signals by explicit wakeup notifications.
The commit also increases the NIC session's buffer size from 128 to 1024
packets.
Issue #4697
...and tighten constness in adjacent code parts.
The VFS-internal synchronization via mutexes is no longer needed because
the access to the VFS is serialized by the VFS client, i.e., the libc.
Issue #4697
Since the lwip library is used only by the vfs_lwip plugin, this patch
moves the Genode-specific support code local to the VFS plugin, easing
further simplification and allowing for a tighter integration with the
VFS plugin.
The recipes for api/lwip and src/vfs_lwip stay separate as both parts
use different licenses.
Issue #4697
This networking scenario is useful for analysing and optimizing the
interplay of the VFS, libc, TCP/IP, and the NIC router. It downloads a
file via fetchurl from lighttpd, both of which are hosted on a virtual
network.
Issue #4697
This patch facilitates the batching of I/O operations in the VFS library
by replacing the implicit wakeup of remote peer (via the traditional
packet-stream interface like 'submit_packet') by explicit wakeup
signalling.
The wakeup signalling is triggered not before the VFS user settles down.
E.g., for libc-based applications, this is the case if the libc goes
idle, waiting for external I/O.
In the case of a busy writer to a non-blocking file descriptor or socket
(e.g., lighttpd), the remote peers are woken up once a write operation
yields an out-count of 0.
The deferring of wakeup signals is accommodated by the new 'Remote_io'
mechanism (vfs/remote_io.h) that is designated to be used by all VFS
plugins that interact with asynchronous Genode services for I/O.
Issue #4697
This test reveals the patters of the batching of consecutive write
operations on a file-system session. It issues 100 writes of one byte
each, which should ideally result in large batches of operations
submitted to the file-system session at once.
The run script performs the write operations through a chain of two VFS
servers, thereby exercising the write batching of both the libc and the
intermediate VFS server.
Issue #4697
By replacing the calls of 'acknowledge_packet' and 'get_packet' with
'try_ack_packet' and 'try_get_packet', we avoid the implicit triggering
of data-flow signals. Instead, the VFS server now relies on explicit
calls of the packet stream's 'wakeup' interface.
Issue #4697
The change of the queue size from 16 to 32 has negligible costs (4 KiB
instead of 2 KiB for the packet-stream queues) while facilitating the
batching of many small consecutive write operations.
Issue #4697
Some signal-heavy scenarios (e.g., libc_integration) produced the
following warning that hinted a data race on signal data in the context
object.
Warning: returning signal with num == 0
The cause was the use of Signal_context::local_submit() in the libc
introduced in
424ed1b79a libc: remove Reconstructible / use local_submit in kernel
in combination with a missing context-mutex aquisition resulting in a
data race on Signal_context::_curr_signal.
Issue #3923
The commit "sculpt_manager: relax nic_drv policy label" introduced the use of
the "label_prefix" attribute instead of "label" for the uplink policy in the
NIC router. However, it missed an appropriate adaption of the lookup of that
attribute when the Sculpt manager has to decide which uplink is used in a
manually managed router config. This caused the uplink to disappear whenever a
user created a manually managed router config. This commit fixes the problem.
Issue #4660Fixes#4695
This prevents errors like the following during parallel builds.
MERGE ld-hw.abi.so
/usr/local/genode/tool/21.05/bin/genode-arm-ld:symbol.map:0: syntax error in VERSION script
